This year, our inspiration comes from a structure that dominates the world around us: networks. From the Internet and routes of public transport to disease transmission and spider webs, networks abundantly surround human-centered and natural activities by providing nodes and links for the exchange of information, ideas or physical beings.
The central concept of our project can be summarized by the greek word “αντιδάνειο” [andiδánio], which, in english, roughly translates to reborrowing, loanword, or repatriated loan. The term loanword is defined as “a word loaned from one language into another and borrowed back (in a modified form and/or meaning) by the first language” [1] and elegantly describes our brainstorming path. By drawing inspiration from neural circuits in the brain, scientists of the previous century devised the first Artificial Neural Network (ANN), forever revolutionizing the data processing domain. Since data processing is crucial to biological procedures as well, we thought that going backwards and designing a bacterial artificial intelligence system based on ANNs would offer new prospects in several fields. So, following in the footsteps of a loanword, we went from Biology to Computer Science and back to Biology.
It is true that cells can process environmental information and proceed to a response based on this very information. In most applications of biological engineering, single cells or cellular systems are deployed for the detection of specific molecules that lead to a specific reaction. Therefore, all these applications are based on the ability of cells to process information. However, one of the current problems in Synthetic Biology is that single cell engineering, and thus information processing, is limited by two factors: firstly, there is the necessary molecular complexity and, secondly, this complexity is accompanied by the metabolic burden of synthetic circuits.
Our idea to create a biological system that implements the perceptron algorithm aims to tackle this problem by being able to process multiple complex information in order to “make a decision” and proceed to a specific action.
The Perceptron is a single layer neural network that is considered as a Linear Binary Classifier. In machine learning, binary classification is a supervised learning algorithm that categorizes novel observations into one of two classes.
The basic parts of a perceptron are the neurons (or nodes), the input values, the weights and biases, the weighted sum, the activation function and the output values. Weights show the strength of the particular node and a bias value allows you to shift the activation function curve up or down. All the inputs x are multiplied with their weights w. All the multiplied values and the biases are added up resulting in the weighted sum. The activation function is used to determine whether the weighted sum is higher than a threshold value. So, the output value represents the class that every novel observation belongs to. Practically, the perceptron is an algorithm that can make a decision between two options and this exact property we took advantage of for the creation of our biological system.
Quorum sensing (QS) constitutes a well-studied mechanism of intercellular communication among bacteria, which enables them to detect alterations in cell population density and regulate their gene expression accordingly. Like hormones in the human body, bacterial exchange of information of this sort is achieved through synthesis of specialized signaling molecules, resulting in coordination of several biochemical processes and communal behaviours, such as sporulation, bioluminescence, virulence, conjugation, competence and biofilm formation.
In the case of Gram-negative bacteria, including E.coli, the signaling molecules -also called autoinducers- belong to the family of acyl-homoserine lactones (AHLs). AHLs are produced throughout the reproductive cycle of bacteria and diffuse through their cell wall, gradually accumulating in the extracellular space as bacteria multiply. Once their concentration in the growth medium reaches a critical point, it becomes thermodynamically unfavourable for AHL molecules to continue to exit the cell. The consequent increase in their intracellular concentration leads to their binding to specific receptors, ultimately activating the expression of key genes [2].
This year, we aim to design and implement the perceptron algorithm in a biological system by taking advantage of the quorum sensing mechanism. With our project, we hope to take bacterial artificial intelligence a step further by building a biological perceptron not in a single cell but on a population scale.
We have engineered two E.coli populations, the senders and the receivers, in order to perform a binary classification task, using the well-characterized
QS system of LuxR/I, that utilizes Acyl-Homoserines Lactones. Senders will evaluate the inputs and communicate their importance to the receiver population
to make a decision. Different synthetic RBSs are introduced to senders to control the expression of the QS molecule (LuxI➡︎OC6), resulting in different OC6
concentrations. The total OC6 produced is received by the receivers, where a genetic circuit with a steep activation response to OC6 is engineered, to
resemble an ON/OFF switch. Finally, mNeonGreen is produced based on the aforementioned biological activation function system and… a decision
is made.
The simplest implementation of a perceptron performs the following steps given a sequence of inputs (x1, x2, …, xn):
Step 1: Weights Assignment: It assigns weights (w1, w2, …wn) to the given inputs.
Step 2: Summation: It calculates the summation Σ(wixi)
Step 3: Activation Function: It feeds the summation to an activation function.
Step 4: Classification: Based on the output of the activation function, it decides whether this pattern belongs to class A (0) or class B (1).
Implementing the perceptron in our system:
Step 1: Weights Assignment: Synthetic RBSs are introduced in the different sender bacteria. The expression of the QS molecule (LuxI->OC6) is controlled by each specific RBS resulting in different quantities of OC6 being produced.
Step 2: Summation: The total OC6 produced is summed and received by the receivers.
Step 3: Activation Function: Our receiver genetic circuit is engineered with a steep activation response to OC6 to resemble an ON/OFF switch.
Step 4: Classification: mNeonGreen is produced based on the biological activation function system and in this way the class is chosen, or it could be said that a decision is made.
The biological sensor our team aims to create, is intending to analyse complex stimuli. The system we propose was therefore originally designed to recognize 3x1 bits patterns. These patterns are created by on-off bits. We consider “on” a subpopulation of senders that is induced with aTc and “off” a senders' subpopulation that is not induced with aTc, and as a result, there will be no lactone production regardless of the RBS. The possible patterns are:
The 166 RBSs that make up the library we have created were divided into three families according to the rate of translation they show when followed by the LuxI gene. The RBSs that have high translation initiation rate are called the “Strong (S)” RBSs, the ones with medium translation initiation rate are called the “Medium (M)” RBSs and the ones with low translation initiation rate are called “Weak (W)”. So by creating different combinations of RBS triads, we expect different amounts of lactone produced.
Thus combining RBSs with different strength in populations that are induced or not, we create patterns that can be provided to the senders as inputs,
which our system will be able to recognize. Theoretically, from each 3x1 bit pattern, we expect a different amount of lactone production when all senders
subpopulations are induced with aTc. Since the strength of each RBS is translated into an amount of lactone produced and finally the subpopulations of
senders are mixed, the position of each bit does not matter but only what RBS each subpopulation has and if the population is induced or not. Furthermore,
it is obvious that the maximum amount of the lactone will be produced in case no. 2 where the subpopulations of the senders all have very strong RBSs.
Respectively, the lower amount of the lactone is expected in case no. 4 where the subpopulations of the senders all have very weak RBSs.
Our proposed implementation of the perceptron weights with the use of synthetic RBS sequences could become the template for the development of
plug-and-play devices that enable the modification of the inputs' weights regarding the desired behaviour and application. We selected to use RBSs
instead of promoters for the pattern recognition because of their versatility and it is quicker to and easier to predict the translation initiation rate
of an RBS instead of direct evolutionize a promoter to have the desired strength. Overall, such a multicellular system has various applications in many
scientific fields, from medicine to environment and the thriving field of biocomputing. It could become the starting point for the development of “smart
biosensors” or “smart drugs” and even perform complex computational tasks using much less energy than the amount needed by a computer, aiming to solve one
of the most popular problems nowadays, the high cost of computational performances.
Adding to previous and current work, our project will provide an alternative point of view in the design of synthetic biological systems. Our proposed implementation of the perceptron weights with the use of synthetic RBS sequences could become the template for the development of plug-and-play devices that enable the modification of the inputs' weights regarding the desired behavior and application. Overall, such a multicellular system has various applications in many scientific fields, from medicine to environment and the thriving field of biocomputing. It could become the starting point for the development of “smart biosensors” or “smart drugs” and even perform complex computational tasks using much less energy than the amount needed by a computer, aiming to solve one of the most popular problems nowadays, the high cost of computational performances.
© 2022 - Content on this site is licensed under a Creative Commons Attribution 4.0 International license