Expanding biological control to bioelectronics with machine learning

Beginning with Galvani’s studies on animal electricity,¹ bioelectronics has merged electronic devices with living systems. Advances in microfabrication and novel materials have now reduced the size of the bioelectronic interface from macroscale contacts capable of stimulating entire muscles to devices that can communicate with individual cells^2,3 and membrane proteins.^4–7 This has blurred the boundaries between bioelectronics, electrophysiology, brain machine interfaces, medical devices, and electroceuticals with much cross-collaboration between the fields.^8,9 Some of the latest bioelectronic devices operate at massive scale, containing thousands of electrodes for sensing with the capability of sub-cellular resolution.^10,11 These sensors and actuators are composed of different functional materials such as inorganic-,^12,13 organic-,^14,15 and nano-materials¹⁶ and share different modes of operation including magnetic,^17,18 electrochemical and redox-based,^19–21 plasmonic,²² and optoelectronic²³ devices. These technologies have applications in interfacing with various levels of biology from the sub-cellular components^24,25 to organs²⁶ and organisms.²⁷ The steady development of bioelectronics has opened opportunities to exploit bioelectronic sensors and actuators to achieve biological control in systems such as in synthetic biology²⁸ and therapeutics.²⁹ 

Nature achieves biological control using feedback wherein sensors and actuators work in unison to maintain a steady state such as in homeostasis or achieve a desired target state such as in sensorimotor movements.³⁰ In this paper, biological control refers to the achievement of an intended and predicted response in a biological system. Broadly speaking, feedback is the feature of using information on system response to past inputs to inform immediate action.³¹ This type of control is referred to as closed-loop control. To this end, efforts toward integrating closed-loop control in biological systems have advanced primarily in synthetic biology.^32–35 Closed-loop control in synthetic biology mimics natural systems such as genetic engineering,^36–38 automated manipulation of the external environment through microfluidics,^36,39 and glucose regulation with an artificial pancreas.⁴⁰ In contrast, relatively few examples of closed-loop control using bioelectronics currently exist.^41,42 In this Perspective, we review closed-loop control approaches in bioelectronics and we propose how using machine learning (ML) can broaden the application to these approaches both in controlling synthetic biology and medical devices (Fig. 1).

Bioelectronics approaches for the control of cellular behavior are often focused on health care applications such as pacemakers, neural implants, and electroceuticals.^43–46 Applications in peripheral neuromodulation has gained the most interest with great potential for closed-loop control to help regain motor control.⁴⁷ The integration of feedback provides an advantage over open-loop control by increasing the efficiency and minimizing the side effects [Fig. 2(a)].

Current methods in closed-loop control for biological systems inside and outside of bioelectronics rely on traditional tools from control theory. For example, another application that has gained similar traction as neural modulation is in diabetes management. Control efforts for an artificial pancreas have been numerous⁴⁸ but generally lie in one of two strategies. The first strategy is a proportional-integral-derivative (PID) controller, which is a common method in the industry and thought to mimic nature [Fig. 2(b)]. The second strategy is a model-predictive controller (MPC), which uses a model to forecast the effects of control on the system [Fig. 2(c)]. MPCs are used to predict the effects of insulin delivery on glucose levels and pick the optimal strategy to achieve the desired glucose level.^47,49

While PID is the commonly adopted controller in practical engineering, its performance degrades when the system is exposed to the non-linearities, uncertainty, and unknown disturbances induced by unwanted phenomena such as variational changes in environmental conditions [Fig. 2(b)]. In diabetes management, this can come from dietary changes or increased physical activity. While researchers have developed methods to improve PID performance by increasing robustness to non-linear systems with uncertainties and process noise,^50–52 the primary hurdle with the PID controller is that conventional approaches to designing the parameters of the PID rely on a time-invariant linear or linearized model of the system.^50–52 This becomes a problem when dynamics become unpredictable or target goals increase in complexity and move away from maintaining homeostasis. On the other hand, an MPC approach can handle nonlinearities and track time varying target outputs; however, it can be sensitive to unmodeled dynamics [Fig. 2(c)]. These approaches are not immune to the inherent properties of nature such as stochasticity⁵³ and unmodeled dynamics.^53,54

The successful applications of control systems theory to biological systems have primarily relied on utilizing a detailed set of mathematical tools to initially model biological systems with precision.⁵⁵ These models require significant a priori knowledge underlying the chemical and genetic mechanisms governing the overall systems.^30,56 Even if this knowledge were to exist, biological systems are adaptive, which means that their response to control and actuation varies with time. As a result, developing closed-loop controllers that can evaluate a signal from a sensor and drive an actuator accordingly is particularly challenging for bioelectronic systems due to the variable conditions that make standard closed-loop control mechanisms unreliable. Here, we argue that methods in ML provide a suitable platform to overcome these challenges but require further development and describe a proof-of-concept example of the ML control of pH and cellular response.

ML-based techniques are suitable when accurate control is required in the absence of a precise mathematical model.⁵⁷ The best-known ML techniques rely on the availability of large datasets a priori and have not been applied to control bioelectronic devices.^58–61 We propose that ML-based techniques that are explored as control solutions outside of biology for cases involving complex non-linear systems are also suitable for closing the loop for bioelectronic systems containing biosensors, biology, and bioelectronic actuators. To this end, tools from control systems theory leveraging ML can be used to learn from new observations for effective real-time operation without data a priori.^62–65 ML-based techniques can be implemented directly [Fig. 3(a)] or indirectly [Fig. 3(b)] to solve complex control problems.^66,67

An example of the ML control of bioelectronics is the real-time ML control of pH in solution recently demonstrated by the authors.^63,68 This approach can be expanded to the control of cell membrane voltage (V_mem) facilitated by pH with an addition of control hierarchy [Fig. 4(a)]. This hierarchy contains the following three levels: a decision maker, a planner, and a low-level controller. The decision maker sets the short-term goals (pH values) that, in turn, affect the long-term goals (V_mem). The planner organizes the activities between those required by the decision maker and the low-level controller that directly interfaces with the bioelectronic actuators and sensors. An example of an activity required by the decision maker is mapping V_mem in response to the change in pH. An example of an activity required by the low-level controller is adjusting the setpoints for a more effective operation of bioelectronics actuators. The low level-controller interfaces directly with the bioelectronic actuator and sensor to provide data for the higher-level controllers.

In this system, pH changing actuators affect the [H⁺] surrounding a cell and shift the cell’s membrane potential [Fig. 4(b)]. This shift is measured with a voltage sensitive fluorophore. The fluorescence intensity of the fluorophore is fed back into the controller and compared with the set value to see whether further actuation is required or not [Fig. 4(b)]. The advantage of using ML for this system is that no assumptions are made on the most suitable control algorithm, which would be impossible to determine a priori given the uncertainty associated with cell response to actuation. In this case, uncertainties and non-linearities of the system arise from elements of the system that occur at different timescales including temporal ionic currents that occur in response to the electric field, the membrane bound ion channel activity, and the expression of stress response factors. Here, a neural network learns how to drive the bioelectronic device to achieve the desired response in the cells despite these uncertainties. This network is robust to variability and noise both in the cell response and the bioelectronic devices themselves.

Nature and life itself revolve around the closed-loop control of biological processes. With bioelectronic devices, we have the opportunity to sense and actuate biological systems and we require powerful control algorithms to do so. These algorithms must be robust to noise, uncertainty, and variability in time and response typical of biological systems. ML-based techniques can provide accurate control overcoming these limitations, and because they can learn and improve with time, they require no a priori knowledge of the system. Here, we have described ML approaches to control that can expand the reach of bioelectronic devices and have shown an example of an application of the ML control of the cell membrane voltage via bioelectronic pH control. We predict that adding ML control to bioelectronics will facilitate the convergence of bioelectronics with synthetic biology²⁸ opening up many new opportunities for both fields and enhance therapeutic applications.⁶⁹ 

J.S. and M.J. contributed equally to this work. All authors wrote and edited this Perspective article.

Data sharing is not applicable to this article as no new data were created or analyzed in this study.

The authors acknowledge funding from the Defense Advanced Research Projects Agency (DARPA), Army Research Office under Cooperative Agreement No. W911NF-18-2-0104 and the Department of Interior under Cooperative Agreement No. D20AC00003. The content of the information does not necessarily reflect the position or the policy of the Government, and no official endorsement should be inferred.