The Use of Machine Learning to Develop Biocompatible Electronics

one month ago by Kristian Nelkovski

Researchers from the University of Chicago, Pritzker School of Molecular Engineering have announced the discovery of new candidate materials for biocompatible electronic devices using machine learning.

What Are Biocompatible Electronic Devices?

Biocompatible electronic devices are a form of implantable electronic circuitry (transistors, sensors, transceivers, and other types of semiconductors) engineered using the unique properties of materials in order to establish communication between a living organism and an outside computer.

Biocompatible electronics have been the subject of interest for many biologists and engineers in the past couple of decades, with the main goal being the creation of human implantable devices for personalised monitoring, diagnosis, analysis and treatment of patients from a wide variety of illnesses and injuries.


A Base for the Future of Biocompatible Electronics 

The most recent materials that are being experimented on for the use of creating bioelectronics are pi-conjugated oligopeptides, which are a type of self-assembling peptide that could potentially become the basis of future biocompatible electronic devices.

The problem with these materials is that it’s difficult to identify the correct molecular sequence in order to engineer the optimal self-assembled nanostructure. Every sequence requires roughly a month of lab testing and there are thousands of potential possibilities to be tested.


Machine learning graphic.

Machine learning tools developed by Professor Ferguson and his colleagues. Image Credit: Kirill Shmilovich et al.


The Problem of Individual Laboratory Materials Testing 

In order to combat this, Associate Professor of Molecular Engineering Andrew Ferguson and his colleagues have developed machine learning instruments to significantly accelerate the process of finding the optimal self-assembling nanostructures from the pi-conjugated oligopeptides.

Using machine learning instruments along with molecular simulation, Professor Ferguson and graduate student Kiril Shmilovich were able to screen and rank 8000 potential oligopeptide candidates, the most promising of which can undergo further laboratory testing and experimentation. All of the candidate oligopeptides were based on one core molecule where each iteration had three symmetrical amino acids on both sides of the molecule changed.


The Bayesian Optimisation or Active Learning Model

The machine learning method used in the study to guide the molecular simulations is called Bayesian optimisation (also known as active learning) which was able to give reliable unbiased data for the model of how the changes of the oligopeptide molecule affected its characteristics.

The model was built by using only 186 oligopeptides which were enough for the simulation to be able to predict the characteristics of the rest of the oligopeptides in the researched peptide family. 

The candidates that showed the most potential from the 8000 ranked peptides were catalogued and handed off to the experimental laboratories of Professor Ferguson’s colleagues in order to undergo further real-life testing outside of the computer simulation.


Organic CMOS logic circuit.

An organic CMOS logic circuit.


Removing Human Bias

Because the Bayesian optimisation method removes human interference within the machine learning simulation, artificial intelligence was able to independently guide the formulation of the oligopeptides instead of the researchers.

This way the simulation was able to find and consider peptide designs that were previously overlooked or even dismissed by scientists. Some of these disregarded peptide designs showed promising characteristics after being tested by the simulation making them candidate materials for engineering self-assembled nanostructures for use in bioelectronics.


The Future of the System

Testing different pi-conjugated cores are planned for the future development of the system. Through machine learning of how different cores and different amino acids change the properties of the peptide, the model of the simulation will be strengthened and equipped to digitally (within the simulation) find and rank new optimal oligopeptides for further laboratory testing.

According to Professor Ferguson, the system can also be used to design proteins and optimise self-assembling colloids for making atomic crystals. With the creation and advancement of these tools, the goal is to create a self-driving laboratory where artificial intelligence can analyse data and make predictions for experiments.

This system is also intended to provide a feedback loop of gathered data to the machine learning algorithm in order to improve the model itself, fully unbiased, without human interference.