Abstract
Wetware computing and organoid intelligence is an emerging research
field at the intersection of electrophysiology and artificial
intelligence. The core concept involves using living neurons to perform
computations, similar to how Artificial Neural Networks (ANNs) are used
today. However, unlike ANNs, where updating digital tensors (weights)
can instantly modify network responses, entirely new methods must be
developed for neural networks using biological neurons. Discovering
these methods is challenging and requires a system capable of
conducting numerous experiments, ideally accessible to researchers
worldwide. For this reason, we developed a hardware and software system
that allows for electrophysiological experiments on an unmatched scale.
The Neuroplatform enables researchers to run experiments on neural
organoids with a lifetime of even more than 100 days. To do so, we
streamlined the experimental process to quickly produce new organoids,
monitor action potentials 24/7, and provide electrical stimulations. We
also designed a microfluidic system that allows for fully automated
medium flow and change, thus reducing the disruptions by physical
interventions in the incubator and ensuring stable environmental
conditions. Over the past three years, the Neuroplatform was utilized
with over 1,000 brain organoids, enabling the collection of more than
18 terabytes of data. A dedicated Application Programming Interface
(API) has been developed to conduct remote research directly via our
Python library or using interactive compute such as Jupyter Notebooks.
In addition to electrophysiological operations, our API also controls
pumps, digital cameras and UV lights for molecule uncaging. This allows
for the execution of complex 24/7 experiments, including closed-loop
strategies and processing using the latest deep learning or
reinforcement learning libraries. Furthermore, the infrastructure
supports entirely remote use. Currently in 2024, the system is freely
available for research purposes, and numerous research groups have
begun using it for their experiments. This article outlines the
system’s architecture and provides specific examples of experiments and
results.
Keywords: wetware computing, organoid intelligence, biocomputing,
synthetic biology, AI, biological neural network, hybrot
1. Introduction
The recent rise in wetware computing and consequently, artificial
biological neural networks (BNNs), comes at a time when Artificial
Neural Networks (ANNs) are more sophisticated than ever.
The latest generation of Large Language Models (LLMs), such as Meta’s
Llama 2 or OpenAI’s GPT-4, fundamentally rely on ANNs.
The recent acceleration of ANN use in everyday life, such as in tools
like ChatGPT or Perplexity combined with the explosion in complexity in
the underlying ANN’s architectures, has had a significant impact on
energy consumption. For instance, training a single LLM like GPT-3, a
precursor to GPT-4, approximately required 10 GWh, which is about 6,000
times the energy a European citizen uses per year. According to a
recent publication the energy consumption projected may increase faster
than linearly ([31]De Vries, 2023). At the same time, the human brain
operates with approximately 86 billion neurons while consuming only
20 W of power ([32]Clark and Sokoloff, 1999). Given these conditions,
the prospect of replacing ANNs running on digital computers with real
BNNs is enticing ([33]Smirnova et al., 2023). In addition to the
substantial energy demands associated with training LLMs, the inference
costs present a similarly pressing concern. Recent disclosures reveal
that platforms like OpenAI generate over 100 billion words daily
through services such as ChatGPT as reported by Sam Altman, the CEO of
OpenAI. When we break down these figures, assuming an average of 1.5
tokens per word—a conservative estimate based on OpenAI’s own tokenizer
data—the energy footprint becomes staggering. Preliminary calculations,
using the LLaMA 65B model (precursor to Llama 2) as a reference point,
suggest energy expenditures ranging from 450 to 600 billion Joules per
day for word generation alone ([34]Samsi et al., 2023). While necessary
for providing AI-driven insights and interactions to millions of users
worldwide, this magnitude of energy use underscores the urgency for
more energy-efficient computing paradigms.
Connecting probes to BNNs is not a new idea. In fact, the field of
multi-unit electrophysiology has an established state of the art
spanning easily over the past 40 years. As a result, there are already
well-documented hardware and methods for performing functional
electrical interfacing and micro-fluidics needed for nutrient delivery
([35]Gross et al., 1977; [36]Pine, 1980; [37]Wagenaar et al., 2005a;
[38]Newman et al., 2013). Some systems are also specifically designed
for brain organoids ([39]Yang et al., 2024). However, their research is
mostly focused on exploring brain biology for biomedical applications
(e.g., mechanisms and potential treatments of neurodegenerative
diseases). The possibility of using these methods for making new
computing hardware has not been extensively explored.
For this reason, there is comparatively less literature on methods that
can be used to reliably program those BNNs in order to perform specific
input–output functions (as this is essential for wetware computing, not
for biomedical applications). To understand what we need for
programming of BNNs, it is helpful to look at analogous problem for
ANNs.
For ANNs, the programming task involves finding the network parameters,
globally denoted as
[MATH: S :MATH]
below, that minimize the difference
[MATH: L :MATH]
computed between expected output
[MATH: E :MATH]
and actual output
[MATH: O :MATH]
, for given inputs
[MATH: I :MATH]
, given the transfer function
[MATH: T :MATH]
of the ANN. This can be written as:
[MATH: L=fOE :MATH]
, with
[MATH: O=TIS :MATH]
where
[MATH: f :MATH]
is typically a function that equals 0 when
[MATH: O=E :MATH]
.
The same equation applies to BNNs. However, the key differences
compared to ANNs include the fact that the network parameters
[MATH: S :MATH]
cannot be individually adjusted in the case of BNNs, and the transfer
function
[MATH: T :MATH]
is both unknown and non-stationary. Therefore, alternative heuristics
must be developed, for instance based on spatiotemporal stimulation
patterns ([40]Bakkum et al., 2008; [41]Kagan et al., 2022; [42]Cai et
al., 2023a,b). Such developments necessitate numerous
electrophysiological experiments, including, for instance, complex
closed-loop algorithms where stimulation is a function of the network’s
prior responses. These experiments can sometimes span days or months.
To facilitate long-term experiments involving a global network of
research groups, we designed an open innovation platform. This platform
enables researchers to remotely perform experiments on a server
interfaced with our hardware. For example, our Neuroplatform enhances
the chances of discovering the abovementioned stimulation heuristics.
It should be noted that, outside of the field of neuroplasticity,
similar open platforms were already proposed in 2023 ([43]O’Leary et
al., 2022; [44]Armer et al., 2023; [45]Elliott et al., 2023; [46]Zhang
et al., 2023). However, to our knowledge, there are no platforms
specifically dedicated to research related to biocomputing.
2. Biological setup
The biological material used in our platform is made of brain spheroids
[also called minibrains ([47]Govindan et al., 2021), brain organoids
([48]Qian et al., 2019), or neurospheres ([49]Brewer and Torricelli,
2007)] developed from Human iPSC-derived Neural Stem Cells (NSCs),
following the protocol of Prof. Roux Lab ([50]Govindan et al., 2021).
Based on the recent guidelines to clarify the nomenclature for defining
3D cellular models of the nervous system ([51]Paşca et al., 2022), we
can call those brain spheroids “forebrain organoids” (FOs). Generation
of brain organoids from NSCs has been already described for both mouse
([52]Ciarpella et al., 2023), and human models ([53]Lee et al., 2020).
Our protocol is based on the following steps: expansion phase of the
NSCs, induction of the 3D structure, differentiation steps (using GDNF
and BDNF), and maturation phase ([54]Figures 1A,[55]B). The [56]Figure
1C is an image of the FO obtained using electronic microscope, it shows
that it is a compact spheroid. The average shape of FOs obtained with
this protocol are spheroids of a diameter around 500 μm ([57]Govindan
et al., 2021). Our experiments show that the FOs obtained can be kept
alive in an orbital shaker for years, as previously demonstrated
([58]Govindan et al., 2021).
Figure 1.
[59]Figure 1
[60]Open in a new tab
FO generation and MEA setup. (A) Protocol used for the generation of
forebrain organoids (FO). Neural progenitors are first thawed, plated
and expanded in T25 flasks. They are then differentiated in P6 dishes
on orbital shakers, and finally manually placed on the MEA. (B)
Representative images showing various stages of FO formation and
differentiation, taken at different time points. The scale bar
represents 250 μm. (C) Image of a whole FO taken with scanning electron
microscope. The scale bar represents 100 μm. (D) Microscope view of the
FO (in white) sitting on the electrodes of the MEA, and the membrane.
The hole in the membrane is not visible on the picture since it is
hidden by the FO. The scale bar represents 500 μm (E) Overview of the
MEA, where the 32 electrodes are visible as 4 sets of 8 electrodes
each. An FO is placed atop of each set of 8 electrodes, visible as a
darker area. For each FO, the 2 circles correspond to a 2.5 mm circular
membrane with a central hole. The scale bar represents 1 mm. (F)
Cross-sectional view of the MEA setup, illustrating the air-liquid
interface. The medium covering the FO is supplied from the medium
chamber through the porous membranes.
Gene expression analysis of mature FOs vs. NSCs showed a marked
upregulation of genes characteristic to neurons, oligodendrocytes and
astrocytes in FOs compare to NSCs. More precisely, FOs expressed genes
typically enriched in the forebrain, such as striatum, sub pallium, and
layer 6 of motor cortex ([61]Govindan et al., 2021). Pathway enrichment
analysis of FOs vs. NSCs demonstrated activation of biological
processes like synaptic activity, neuron differentiation and
neurotransmitter release ([62]Govindan et al., 2021).
At the age of 12 weeks, FOs contain a high number of ramified neurons
([63]Govindan et al., 2021), and they are mature enough to be
transferred to the electrophysiological measurement system ([64]Figure
1A). In this setup, they have a life expectancy of several months, even
with 24/7 experiments that include hours of electrical stimulations.
This setup has a quick turnaround with occasional downtime – about 1 h
– during organoid replacements. Therefore, the platform maintains a
high availability for experiments.
3. Hardware architecture
3.1. Introduction
The remotely accessible hardware includes all the systems which are
required to preserve homeostasis, monitor environmental parameters and
perform electrophysiological experiments. These systems can be
controlled interactively using our custom Graphical User Interface
(GUI) or via Python scripts. All data is stored in a time-series
database (InfluxDB), which can be accessed either via a GUI or via
Python scripts. The users typically connect to the system using the
Remote Desktop Protocol (RDP).
The platform is composed of several sub-systems, which can be accessed
remotely via API calls over the internet, typically through Python.
3.2. Multi-Electrode Array (MEA)
Our current platform features 4 MEAs. The MEAs were designed by Prof.
Roux’s Lab form Haute Ecole du Paysage, d’Ingénierie et d’Architecture
(HEPIA) and are described in [65]Wertenbroek et al. (2021). Each MEA
can accommodate 4 organoids, with 8 electrodes per organoid ([66]Figure
1E).
The MEA setup utilizes an Air-Liquid-Interface (ALI) approach
([67]Stoppini et al., 1991), in which the organoids are directly placed
on electrodes located atop of a permeable membrane ([68]Figure 1D),
with the medium flowing beneath this membrane in a 170 μL chamber. As a
result, a thin layer of medium, created by surface tension, separates
the upper side of the organoids from the humidified incubator air. This
arrangement is further protected by a lid partially covering the MEA
([69]Figure 1F). This ALI method enables a higher throughput and higher
stability compared to submerged approaches, since no dedicated coating
is required, and it is less prone to have the organoids detaching from
the electrodes.
3.3. Electrophysiological stimulation and recording system
The electrodes in our system enable both stimulation and recording. The
respective digital-to-analog and analog-to-digital conversions are
performed by Intan RHS 32 headstages. Stimulations are executed using a
current controller that ranges from 10 nA to 2.5 mA, and recordings are
obtained by measuring the voltage on each electrode at a 30 kHz
sampling frequency with a 16 bits resolution giving an accuracy of
0.15 μV. The headstages are connected to an Intan RHS controller, which
in turn is connected to a computer via a USB port. The [70]Figure 2A
shows the electrical activity recorded for each of the 32 electrodes.
It can be noticed that the recorded activity is different between each
electrode. This difference comes from the facts that each set of 8
electrodes records a different FO and that for a given FO, electrodes
record at a different location. This display is refreshed in real-time
and also available 24/7 on our website at the URL
[71]https://finalspark.com/live/. We compared the recording
characteristics of this ALI setup to MCS MEA (60MEA200/30iR-Ti)
monitoring a submerged FO, using the exact same Intan system for
voltage conversion. The overlays of an action potential recorded,
respectively, with the ALI and submerged versions are shown in
[72]Figures 2C,[73]D and show similar signal characteristics.
Figure 2.
[74]Figure 2
[75]Open in a new tab
Recording system and user interface. (A) Electrical activity measured
in μV over one second for each of the 32 electrodes. Each set of 8
electrodes records a different FO. (B) Graphical User Interface for
manually controlling each of the microfluidic pumps. (C) Overlays of FO
action potential recorded by the ALI system of the Neuroplatform. (D)
Overlays of FO action potentials recorded with an MCS system. (E)
Fluctuations of the flowrate of the medium within the microfluidic
system, illustrating the cyclic variations induced by the peristaltic
pump operating at 1 round per minute with 10 cams. (F) Temporal
variations of the red component of the medium color, triggered by a
sudden change in medium acidity, resulting in phenol red color change.
3.4. Micro-fluidics
To sustain the life of the organoids on the MEA, Neuronal Medium (NM)
needs to be constantly supplied. Our Neuroplatform is equipped with a
closed-loop microfluidic system that allows for a 24/7 medium supply.
The medium is circulating at a rate of 15uL/min. The medium flow rate
is controlled by a BT-100 2 J peristaltic pump and is continuously
adjusted according to needs, for instance during experimental runs. The
peristaltic pump is connected to the PC-control software using an RS485
interface, for programmed (i.e., in Python) or manual operations
([76]Figure 2B). Additionally, [77]Figure 3A depicts this microfluidic
closed-loop circuit.
Figure 3.
[78]Figure 3
[79]Open in a new tab
Microfluidics. (A) Microfluidic system illustrating the continuously
operating primary system, which ensures constant flow in the medium
chamber, and the secondary system responsible for medium replacing
every 48 h. (B) Side view of the assembly, featuring the camera and the
MEA. The entire assembly is enclosed with aluminum foil to ensure the
lowest possible noise level. (C) Front view of the assembly, showing
the intake and outtake of the microfluidic system, as well as the LED
used during image capture.
The microfluidic circuit is made of 0.8 mm (inside diameter, ID)
tubing. Continuous monitoring of the microfluidic circuit and flow rate
is achieved by using Fluigent flow-rate sensors, which connect to the
Neuroplatform control center via USB. Data related to medium flow rate
is stored in a database for later access. [80]Figure 2E shows the
cyclic variations in flow induced by the cams of the peristaltic pump.
A secondary microfluidic system is used to replace the medium in the
closed-loop with fresh medium every 24 h, a process illustrated in
[81]Figure 3A. This replacement is fully automated through a Python
script and performed in the following consecutive steps:
1. Set the rotary valve to select the path from the reservoir F50 to
the syringe pump
2. Pump 2 mL of old medium using the syringe pump
3. Set the rotary valve to select the path from the syringe pump to
the waste F50
4. Push 2 mL of old medium to the waste using the syringe pump
5. Set the rotary valve to select the path from the new medium in the
F50 in the fridge to the syringe pump
6. Pump 2 mL of fresh medium using the syringe pump
7. Set the rotary valve to select the path from the syringe pump to
the reservoir F50
8. Push 2 mL of fresh medium using the syringe pump
3.5. Cameras
Each MEA is equipped with a 12.3-megapixel camera that can be
controlled interactively or programmatically (i.e., through a Raspberry
Pi) for still image capture or video recording. The camera is
positioned below the MEA, while illumination is provided by a remotely
controlled LED situated above the MEA. [82]Figures 3B,[83]C illustrate
this assembly (the aluminum wrapping is used in order to minimize the
noise). This setup is particularly useful for detecting various
changes, such as cell necrosis, possible organoid displacement caused
by microfluidics, variations in medium acidity (using color analysis
since our medium contains Phenol red), contamination, neuromelanin
production (which can happen when uncaging dopamine), overflows (where
the medium inadvertently fills the chamber above the membrane), or
bubbles in the medium. For the latter two events, dedicated algorithms
automatically detect these issues and trigger an alert to the on-site
operator.
Changes of acidity, for example, can be detected by measuring the
average color over a pre-defined window. [84]Figure 2F shows the
evolution of the medium’s red color component, with data points
recorded hourly. The noticeable sudden drop is attributed to the
pumping of medium with a slightly different acidity. This change in
acidity results in a color alteration of the phenol red present in the
medium.
3.6. UV light controlled uncaging
It is also possible to release molecules at specific timings using a
process called uncaging. In this method, a specific wavelength of light
is employed to break open a molecular “cage” that contains a
neuroactive molecule, such as Glutamate, NMDA or Dopamine. A fiber
optic of 1,500 μm core diameter and a numerical aperture of 0.5 is used
to direct light in the medium within the MEA chamber. The current
system, Prizmatix Silver-LED, operates at 365 nm with an optical power
of 260 mW. The uncaging system is fully integrated into the
Neuroplatform and can be programmatically controlled during experiment
runs via our API (see section 5.3).
3.7. Environmental measurements
The environmental conditions are monitored within two incubators. In
both incubators, the following parameters are recorded: CO2, O2
concentrations, humidity, atmospheric pressure and temperature.
Door-opening events are also logged since they have a major impact on
measurements. The primary purpose of this monitoring is to ensure that
experiments are performed in stable and reproducible environmental
conditions.
All these parameters are displayed in real-time in a graphic interface
showing both instant values as well as variations versus time of noise
and flowrates ([85]Figure 4A).
Figure 4.
[86]Figure 4
[87]Open in a new tab
Graphic user interface to monitor critical parameters in the
incubators. (A) Graphical User Interface displaying critical
environmental conditions for the incubator 1, where
electrophysiological experiments are performed, as well as the
incubator 2, where FO are maintained on an orbital shaker. (B) The
display shows environmental data for incubator 1 for specific time
periods, extracted from the database, with door opening events
displayed as dashed line. Noise, Temperature, humidity and pressure are
indicated by different colored lines. The units of each measurement are
normalized between 0 and 1 for the selected time interval.
Incubator 1 houses the MEAs and the organoids used for
electrophysiological experiments. In addition to the mentioned
parameters, flowmeters are also utilized to report the actual flow rate
of the microfluidic for each MEA, as depicted in the graph labelled
“Pump” in [88]Figure 4A. The system’s state is indirectly monitored
through the noise level of each MEA, as shown in the graph labelled
“Noise Intan” in [89]Figure 4A. The noise level is calculated based on
the standard deviation of the electrical signals recorded by the
electrodes over a 30 ms period.
Incubator 2 houses the organoids which are kept in orbital shakers.
Piezoelectric gyroscopes are used to measure the actual rotation speed
of the orbital shakers.
Since all the data is logged in the database, it is also possible to
access the historical measurements through a dedicated GUI ([90]Figure
4B).
4. Software
4.1. General architecture
The core of the system relies on a computational notebook which
provides access to 3 resources ([91]Figure 5A):
Figure 5.
[92]Figure 5
[93]Open in a new tab
Software setup and electrical stimulation. (A) General architecture of
the Neuroplatform. The Jupyter Notebook serves as the main controller,
enabling initiation and reading of spikes, configuration stimulation
signals and access to database via, e.g., Python (B) Parameters of the
stimulation current: settings optimally these parameters can elicit
spikes. Through the Python API, parameters that can be adjusted for the
bi-phasic stimulation signals include the duration (D1) and amplitude
(A1) of the positive current phase, and, respectively, D2 and A2 for
the negative current phase. Additionally, the polarity of the biphasic
signal can be reversed to start with a negative current.
* 1. A database where all the information regarding the Neuroplatform
system is stored
* 2. The Intan software running on a dedicated PC, which is used for:
* Recording the number of detected spikes in a 200 ms time window
* Setting stimulation parameters
* 3. A Raspberry Pi for triggering current stimulation according to
stimulation parameters
4.2. Database
The Neuroplatform records monitored data 24/7 using InfluxDB, a
database designed for time series. Other options are also available.
This database contains all the data coming from the hardware listed in
Section 3.
The electrical activity of the neurons is also recorded 24/7 at a
sampling rate of 30 kHz. To minimize the volume of stored data, we
designed a dedicated process that focuses on significant events, such
as threshold crossings that are likely to be due to action potentials
(spikes). The following pseudo code illustrates the implemented
approach:
* - Each 1-min write buffer to database
* - Each 33 μs
* - For each electrode
* - If, at time t, the voltage exceeds a threshold T
* - Store (in buffer) 3 ms of data [t-1 ms, t + 2 ms]
* - Each 3 s update T
Additionally, a timestamp corresponding to each detected event is also
stored in the database, along with the maximum value of voltage during
the 3 ms spike waveform recording.
The threshold T is computed directly from voltage values sampled each
33 μs, according to the following formula:
[MATH:
T=6∗Mdn<
mfenced open="(" close=")">σi :MATH]
Where
[MATH: σi :MATH]
is the standard deviation computed over a set i of 30 ms consecutive
voltage values, and
[MATH: Mdn :MATH]
represents the median function computed over 101 consecutive
[MATH: σi :MATH]
values. The use of the median reduces the sensitivity to outliers,
which is typically caused by action potentials. In our current setup, a
multiplier of 6 on the median has proven to be a good compromise for
achieving reliable spike detection.
Besides electric tension data, the number spikes recorded per minute is
also computed and stored in the database every minute by a batch
process.
4.3. Recording electrical activity
As previously discussed, the communication among neurons is captured by
the MEA and converted into a voltage signal sampled at 30 kHz. The
Neuroplatform offers two basic access modes to the recorded neural
activity:
1. Raw: raw sampling values.
2. Optimized: waveforms of the raw signal near neuronal spikes,
available directly from the database.
In addition to the aforementioned features, the Neuroplatform offers
even more advanced methods. For instance, it includes counting spikes
over a fixed time period of 200 ms following stimulation, with a 10 ms
delay suppressing the stimulation artifact.
From a technical perspective, accessing the number of spikes can be
accomplished in two different ways:
* - Retrieving the number of spikes per minute from the database
* - Through direct communication with the PC managing the Intan
controller for spike count
The second approach is required when the stimulation protocol demands
real-time responsiveness. This is typically the case for certain
closed-loop strategies. For instance, closed-loop stimulation
strategies have been deployed in primary cortical cultures for
effective burst control ([94]Wagenaar et al., 2005a,[95]b) and for
goal-directed learning ([96]Samsi et al., 2023).
4.4. Syntax for stimulations
Programmatically stimulating the FO on the Neuroplatform is
accomplished by sending an electrical current to the MEA electrodes.
The electrical current profile can be parameterized in a variety of
ways, which is partly shown in [97]Figure 5B. These parameters and
controls include:
* - Basic shape of stimulation signal:
* o Bi-phasic
* o Bi-phasic with interphase delay
* o Tri-phasic
* - Stimulation duration and intensity:
* o Positive (A1) and negative (A2) electrical current intensity
(typical 1uA, ranging from 0.1uA to 20uA)
* o Duration of positive (D1) and negative (D2) stimulation currents
* - Stimulation triggers
* o Single start
* o Table with collection of start triggers
* o Pulse trains
* - MEA electrodes
send_stim_param (electrodes, params)
5. Examples of electrophysiological experiments
To demonstrate the effectiveness of the Neuroplatform, the following
sections will provide an overview of several experiments conducted on
the Neuroplatform at FinalSpark’s Laboratories in Vevey, Switzerland.
5.1. Modification of spontaneous activity
The spontaneous electrical activity of the FO can be represented by the
concept of “Center of Activity” (CA) ([98]Bakkum et al., 2008) which is
defined as a virtual position
[MATH: C :MATH]
on the MEA described by:
[MATH: C=∑k=18Fk⋅XkYk∑k=18Fk
:MATH]
Where
[MATH: XkYk :MATH]
define the spatial position of the 8 electrodes and
[MATH: Fk :MATH]
is the number of spontaneous spikes detected. The interest of the
concept of CA is that its position provides statistical information
about the average location of the activity over the surface of the FO.
The ability to change the position of the CA is interesting because it
also shows the ability to memorize information in the state of the FO.
The coordinates of the CA can be modified using a high frequency
stimulation. In the following experiment we use the following protocol:
* 1) Compute the CA using the number of detected spikes over 500 ms
* 2) Goto 1,100x
* 3) Perform a 20 Hz stimulation during 500 ms using a bi-phasic
current (negative first) of 2 μA of 200 μS, for both phases, on one
electrode
* 4) Wait 1 s
* 5) Goto 5,100x
[99]Figure 6A displays the 100 measured positions of the CA
corresponding to the spontaneous activity before the 20 Hz stimulation
in blue, and after the high-frequency stimulation in red (the average
position is indicated by a cross). A close-up is shown in [100]Figure
6B. The timestamps of the spontaneous activity, before and after
stimulation, are presented in [101]Figures 6C,[102]D, respectively.
Each graph shows one example of the 100 records of 500 ms used to
compute the CA location (showing a decrease of spontaneous firing
activity of electrodes 3, 4 and 6). A noticeable shift in the average
position (shown by a cross) of the CA can be observed before and after
the high-frequency stimulation (as seen in [103]Figure 6A), indicating
a change of state of the biological network. A classifier based on a
simple logistic regression is employed to predict if the network has
received the 20 Hz stimulation. In this particular experiment, the
classification accuracy, computed from the confusion matrix, is 95.5%.
Figure 6.
[104]Figure 6
[105]Open in a new tab
Center of activity modification. (A) Graph showing the 2D layout of the
8 electrodes, the X and Y axis are normalized units showing the spatial
coordinates of the electrodes. All electrodes can be used for both
stimulation and reading. A 20 Hz stimulation signal is applied to
electrode 6. The 100 blue circles represent the positions of the Center
of Activity (CA) before 20 Hz stimulation, while the 100 red circles
indicate the positions after the stimulation. The cross mark the
average position. (B) A closer look at the two groups of CA. (C)
Timestamps depicting the spontaneous activity over 500 ms for each of
the 8 electrodes before the high-frequency stimulation. (D) Spontaneous
activity observed after the high-frequency stimulation, showing a lower
activity of electrodes 6, 4 and 3, compared to (C).
The Neuroplatform allows users to perform both the experimental part
(including stimulation and reading operations) and the visualization of
the CA displacement within the same Python source code. The 500 ms
20 Hz signal is generated directly by the Python source code shown
below. The first trigger.send instruction sends the trigger for the
stimulation on a specific electrode and time.sleep pauses the execution
for 50 ms. Inline graphic
Despite the common perception of Python as being less than ideal for
real-time signal processing due to its inherent latency, our empirical
data reveals a time accuracy of under 1 ms (on an Intel Xeon CPU
E5-2690 v2 @ 3.00GHz), a level of precision that is satisfactory for
the generation of tetanic signals.
5.2. Optimization of stimulation parameters
In this example, the objective is to identify the set of stimulation
parameters that can elicit the maximum number of action potentials
within 200 ms after a stimulation.
Depending on the FOs, their composition, and maturity, only specific
combinations of electrodes and parameters can elicit spikes. In our
experiment, we use an 8-electrode MEA and cycle through several
stimulation signal parameters as shown in [106]Figure 7A. Consequently,
we need to test a total of 342 different parameter-electrode
combinations. The following pseudo code illustrates the Python script
used in this experiment.
Figure 7.
[107]Figure 7
[108]Open in a new tab
Neural activity stimulation and dopamine uncaging. (A) Graph depicting
the number of spikes recorded over 250 ms. The spike counts in orange
were measured following a stimulation, while those in blue were
measured during periods without stimulation. For clarity in
visualization, a small bar is displayed even when no spikes are
detected. (B) Diagram illustrating the different steps involved in the
closed-loop uncaging process of dopamine, which is repeated 240 times.
(C) Timestamps of action potentials from the 8 electrodes before and
after stimulation (shown as red line), showcasing the elicited spikes.
(D) Graph displaying the number of elicited spikes over the 240 steps
of the closed-loop (in blue) alongside the activation events of the UV
light source (red).
* 1) For each set of stimulation parameters
* 2) For each stimulation electrode
* 3) For each recording electrode
* 4) During 15 s, every 250 ms
* 5) Decide between stimulating, or recording spontaneous activity,
with a 50% probability
* 6) Record number of spikes during 200 ms
The aim of probabilistic stimulation and no stimulation in step 5 is to
evaluate the difference between elicited and spontaneous spikes in a
way that ensures there is no bias.
The bar chart in [109]Figure 7A displays a segment of the experimental
results. It shows a 15-s recording from a single electrode,
corresponding to one execution of step 4 in the pseudo code above. Each
bar represents the spike count during a 200 ms period, repeated every
250 ms. The orange bars in this plot are the result of the parameters
selected in step 1 of the pseudo code. The blue bars represent
no-stimulation periods, thus corresponding to the spontaneous activity
of the neurons.
From [110]Figure 7A, we can see that this particular combination of
electrode and parameters reliably elicits responses.
In practice, the Python script can also be used to automatically
display the 342 graphs similar to [111]Figure 7A, allowing the operator
to select the optimal set of parameters. Additionally, it can compute a
scalar metric to characterize the “efficiency” of the parameters, and
automatically identify the optimal parameters.
An example of a parameter maximization metric is given in the equation
below. Let us denote
[MATH: μr :MATH]
and
[MATH: μs :MATH]
the average number of spikes recorded spontaneously or after a
stimulation, respectively, and
[MATH: σr :MATH]
and
[MATH: σs :MATH]
as their standard deviations. The following metric is used:
[MATH: m=μr−μ<
/mi>smaxσrσs :MATH]
The set of parameters that maximize this metric can then be utilized to
perform other experiments requiring elicited spikes, such as
investigating the effect of pharmacological agents on a biological
network’s ability to react quickly to stimulation.
5.3. UV light-induced uncaging of molecules
‘Uncaging’ is a pivotal technique in cellular biology, enabling the
precise control of molecular interactions within cells ([112]Gienger et
al., 2020). It involves the use of photolabile caged compounds that are
activated by specific light wavelengths, releasing bioactive molecules
in a targeted and timely manner. This method is particularly valuable
for studying dynamic processes in neural networks and intracellular
signaling, offering real-time insights into complex biological
mechanisms.
Our Neuroplatform is equipped with all necessary components to perform
uncaging. In this example, we investigate closed-loop stimulation,
where dopamine is used to reward the network when more spikes are
elicited by the same stimulation. The release of the dopamine is
achieved through the uncaging of CNV-dopamine using the UV system
described in section 3.6.
[113]Figure 7B shows the flow chart of the closed-loop uncaging
process. The optimal stimulation parameters are first found using the
technique shown in 5.2 (in this case, a current of 4uA, biphasic with
100uS per phase), which is sent successively to each of the 8
electrodes with a delay of 10 ms between each electrode.
[114]Figure 7C shows the response timestamps of the 8 electrodes for a
period of 1,200 ms, 600 ms before and after the stimulation. The
stimulation event is indicated by the vertical red line. It is
interesting to observe that in this particular case, most of the
elicited spikes originate from 2 electrodes, specifically electrode 112
and electrode 119.
The Python source code implementing the closed-loop process illustrated
in [115]Figure 7B is provided below. We would like to highlight here
how concise the code is. With only 13 lines of code, the entire
closed-loop process has been implemented. Inline graphic
The graph in [116]Figure 7D shows the variation in the number of spikes
elicited during the execution of the script above across 5 h. A general
increase in the number of elicited spikes can be observed. However, it
is obviously not possible to establish causality between the
closed-loop strategy and the observed increase with this single
experiment alone. The primary purpose of this closed-loop experiment is
to demonstrate the flexibility offered by the Neuroplatform.
6. External users of the Neuroplatform
Access to the Neuroplatform is freely available for research purposes.
For researchers lacking lab infrastructure, the Neuroplatform provides
the capability to conduct real-time experiments on biological networks.
Additionally, it allows others to replicate results obtained in their
own lab. The database is shared between all research groups, however
the Python scripts and Jupyter Notebooks are in private sections.
In 2023, 36 academic groups proposed research projects, of which 8 were
selected. At the time of writing, 4 of these have already yielded some
results:
* University Côte d’Azure, CNRS, NeuroMod Institute and Laboratoire
JA Dieudonné: investigates the functional connectivity of FO and
how electrical stimulation can modify it.
* University of Michigan, investigates stimulation protocols that
induce global changes in electrical activity of a FO.
* Free University of Berlin, investigates stimulation protocols that
induce changes in the electrical activity of a FO. Additionally,
this research employs machine learning tools to extract information
from neural firing patterns and to develop well-conditioned
responses. Moreover, it utilizes both shallow and deep
reinforcement learning techniques to identify optimal training
strategies, aiming to elicit reproducible behaviors in the FO.
* University of Exeter, Department of Mathematics and Statistics,
Living Systems Institute, investigates storing and retrieving of
spatiotemporal spiking patterns, using closed-loop experiments that
combine mathematical models of synaptic communication with the
Neuroplatform.
* Lancaster University Leipzig and University of York: characterizes
computational properties of FOs under the reservoir computing
model, with a view to building low-power environmental sensors.
* Oxford Brookes University, School of Engineering, Computing and
Mathematics: investigating the properties of emerging dynamics and
criticality within neural organizations using the FOs.
* University of Bath, ART-AI, IAH: using the free energy principle
and active inference to study the learning capabilities of neurons,
embodied in a virtual environment.
* University of Bristol: stimulating of FOs based on data gathered
from an artificial tactile sensor. Use machine learning techniques
to interpret the FO’s output, investigating their ability to
process real-world data.
7. Discussion and conclusion
The Neuroplatform has now been operational 24/7 for the past 4 years.
During this time, the organoids on the MEA have been replaced over 250
times. Considering that we place at least 4 organoids per MEA, and
change all the organoids simultaneously, this amounts to testing over
1,000 organoids. Initially, their lifetime was only a few hours, but
various improvements, especially related to the microfluidics setup,
have extended this to up to 100 days in best cases. It is important to
note that the spontaneous activity of the organoids can vary over their
lifetime, a factor that must be taken into consideration when
conducting experiments ([117]Wagenaar et al., 2006). Additionally, we
observed that the minimum current required to elicit spikes, computed
using the method described in section 5.2, is increasing over the
lifetime of the organoid. This phenomenon may be linked to an impedance
increase caused by glial encapsulation ([118]Salatino et al., 2017).
The 24/7 recording strategy as described in section 4.2, results in the
constant growth of the database. As of this writing, its size has
reached 18 terabytes. This volume encompasses the recording of over 20
billion individual action potentials, each sampled at a 30 kHz
resolution for 3 ms. This extensive dataset is significant not only due
to its size but also because it was all recorded in a similar in-vitro
environment, as described in section 3.2. We are eager to share this
data with any interested research group.
8. Future extensions
In the future, we plan to extend the capabilities of our platform to
manage a broader range of experimental protocols relevant to wetware
computing. For example, we aim to enable a remote control over the
injection of specific molecules into the medium, facilitating remote
experiments that involve pharmacological manipulation of neuronal
activity. This expansion will provide additional degrees of freedom for
the automatic optimization of parameters influencing neuroplasticity.
Currently, as detailed in Chapter 2, only one differentiation protocol
is used for generating organoids. We plan to introduce additional types
of organoid generation protocols soon, with the aim of exploring a
broader range of possibilities.
Although 32 research groups requested to access to the Neuroplatform,
our current infrastructure only allows us to accommodate 7 groups,
considering our own research needs as well. We are in the process of
scaling-up the AC/DC hardware system to support more users
simultaneously. Additionally, we are currently limited to executing
close-loop algorithms for neuroplasticity on one single FO, as these
algorithms require sending in real-time adapted simulation signals to
each FO. Our software is being updated to run closed-loops in parallel
on up to 32 FO.
9. Methods
9.1. Brain organoid generation
Human forebrain organoids were originated as described in [119]Govindan
et al. (2021). Briefly, Human Neural Stem Cells derived from the human
induced pluripotent stem (hiPS) cell line (ThermoFisher), were plated
in flasks coated with CellStart (Fisher Scientific) and amplified in
Stempro NSC SFM kit (ThermoFischer) complete medium: KnockOut
D-MEM/F12, 2 mM of GlutaMAX, 2% of StemPro Neural supplement, 20 ng/mL
of Human FGF-basic (FGF-2/bFGF) Recombinant Protein, and 20 ng/mL of
EGF Recombinant Human Protein (Fisher Scientific). Cells were then
detached with StemPro^™ Accutase (Gibco) and plated in p6 at the
concentration of 250,000 cells/well. The plates were sealed with
breathable adhesive paper and leads, placed on an orbital shaker at
80 rpm, and culture for 7 days at 37°C 5% CO2. After one week the newly
formed spheroids were put in differentiation medium I (Diff I),
containing DMEM/F-12, GlutaMAX^™ supplement (Gibco), 2% BSA, 1X of
Stempro® hESC Supplement, 20 ng/mL of BDNF Recombinant Human Protein
(Invitrogen), 20 ng/mL of GDNF Recombinant Human Protein (Gibco),
100 mM of N6,2′-O-Dibutyryladenosine 3′,5′-cyclic monophosphate sodium
salt, and 20 mM of 2-Phospho-L-ascorbic acid trisodium salt. After one
week, brain spheroids were put in differentiation medium II (Diff II)
made of 50% of Diff I and 50% of Neurobasal Plus (Invitrogen). After
3 weeks of culture in Diff II, brain organoids were plated in
Neurobasal Plus and kept in the orbital shaker until the transfer on
the MEA. Medium was change once per week.
9.2. Electron microscopy analysis of FOs
Mature FOs were fixed in 2.5% Glutaraldehyde in 0.1 M phosphate buffer
pH 7.4, at RT. After 24 h the samples were processed as described in
[120]Cakir et al. (2019) at the Electron Microscopy Facility of
University of Lausanne. The whole FO images were acquired with Quanta
FEG 250 Scanning Electron Microscope.
9.3. Transfer of FOs on MEA
MEA connected with the microfluid system was moved from the incubator
and placed on a 12.3-megapixel camera system (with an optical lens of
16 mm of focal, giving a magnification power of 21x) inside the cell
culture hood. The lid was removed to access the top of the liquid/air
interface. Sterile Hydrophilic PTFE MEMBRANE Hole ‘confetti’ (diameter
2.5 mm, diameter of the hole 0.7 mm) (HEPIA) were positioned on top of
each electrode and left there 2 min to absorb the medium. FOs were
collected from the plate using wide bore pipette tips (Axygen) and
placed in the middle of confetti, in a 10 μL drop of medium. The
position of the organoids was adjusted with the help of sterile
forceps. After all the organoids were put on place, the chamber was
covered with the plate sealer Greiner Bio-One^™ BREATHseal^™ Sealer
(Fisher Scientific), and with the MEA lid. MEA containing the organoids
were placed immediately back in the cell incubator and were ready to be
used for recording and stimulation. A similar procedure was used for
the positioning of organoids on MCS MEA (60MEA200/30iR-Ti). In this
case the Hydrophilic PTFE MEMBRANE was not used and organoids were
directly laid on the electrodes in a 30 μL drop of medium. Recording of
organoid activity was performed immediately afterwards.
9.4. System design and assembly
Cell culture media was stored in a 50 mL Falcon tube with a multi-port
delivery cap (ElveFlow) and stored at 4°C. Each reservoir delivery cap
contained a single 0.8 mm ID × 1.6 mm OD PTFE tubing (Darwin
Microfluidics), sealed by a two-piece PFA Fittings and ferrule threaded
adapter (IDEX), extending from the bottom of the reservoir to an inlet
port on the 4-port valve head of the RVM Rotary Valve (Advance
Microfluidics SA). Sterile air is permitted to refill the reservoir
through a 0.22-μm filter (Milian) fixed to the cap to compensate for
syringe pump medium withdrawal. A similar PTFE tubing and PFA Fittings
and adapters were used to connect the syringe pump to the 4-port valve
head of the RVM Rotary Valve (Advance Microfluidics SA). Each PTFE
tubing coming from the distribution valve connects with a 50 mL falcon
tube inside the cell culture incubator (Binder) and to a borosilicate
glass bottle (Milian) to collect discarded cell culture medium.
A secondary microfluid system made of 0.8 mm ID × 1.6 mm OD PTFE
tubing, were used to connect each 50 mL falcon tube inside the cell
culture incubator with its own MEA (HEPIA). The connection was through
a precise peristaltic pump BT100-2 J (Darwin Microfluidics) containing
10 rollers. A compute module (Raspberry Pi 4) controlled the
peristaltic pump and the Rotary Valve, through a custom application
program interface (API), using RS485 interface and RS-232 interface,
respectively. A Fluigent flow-rate sensor connected via USB to the
Raspberry Pi 4 allowed the monitoring of the flow rate inside the
microfluidic system between the peristaltic pump and the MEA. Python
was used to develop the software required to carry out automation
protocols.
9.5. Uncaging of dopamine
Carboxynitroveratryl (CNV)-caged dopamine (Tocris Bioscience) was
dissolved in Neurobasal Plus at the concentration of 1 mM, and injected
in the fluidic system. After 3 h from the injection, the uncaging
experiment started as described in paragraph 5.3. UV Silver-LED
fiber-coupled LED (Prizmatix) was used to uncage the dopamine at the
wavelength of 365 nm for 800 ms each time.
Data availability statement
The raw data supporting the conclusions of this article will be made
available by the authors, without undue reservation.
Ethics statement
Ethical approval was not required for the studies on humans in
accordance with the local legislation and institutional requirements
because only commercially available established cell lines were used.
Author contributions
FJ: Writing – original draft, Writing – review & editing. MK: Writing –
original draft, Writing – review & editing. J-MC: Writing – original
draft, Writing – review & editing. FB: Writing – original draft,
Writing – review & editing. EK: Writing – original draft, Writing –
review & editing.
Acknowledgments