Dec 2, 2009
Neural networks differentiate bacteria
Researchers at Oak Ridge National Laboratory and Clemson University, both US, have come up with an automatic recognition process for differentiating M. lysodeikticus from P. fluorescens based on the electromechanical properties of the bacteria. Electromechanical signatures of bacteria and substrate were collected over a selected band of frequencies with ~50 nm spatial resolution. Neural network analysis of the data allowed the team to perform automatic attribution of each type of bacteria and the substrate (click on the image below for more detail).
Identification of bacteria is an important task, but the process can involve lengthy and laborious sample preparation techniques. To speed up and simplify the exercise, the group proposed a relatively fast and easy method for recognizing bacteria.
The method exploits the ability of bacteria to change configuration on application of an electric field. The magnitude of the change in shape of the bacteria strongly depends on many factors including the stiffness of bacteria and the composition of the bacteria's outer shell.
"We tether bacteria with an electrical signal of different frequencies and record how the structure twitches", Maxim Nikiforov told nanotechweb.org. "The recorded information is then processed using artificial intelligence methodologies." The end result is a recognition map, where a computer is able to identify the two different types of bacteria and the substrate on which these bacteria were grown.
An atomic force microscope was used to perform the measurements. In this technique, a very sharp probe (~10 nm contact radius) applies a force to the sample in a controllable fashion. In order to measure electro-mechanical coupling, an alternating voltage was applied to the probe.
The characteristic features of the collected frequency spectra were determined using principal component analysis and the most significant features (the information contained in first five principal components) were then used to train the neural network and perform recognition.
The methodology can be used on any spatially resolved spectroscopic data.
Dubbed "Functional Recognition Imaging", the technique has a wide range of applications. "Given the ubiquity of electromechanical coupling in biological systems – from ion channels to membrane flexoelectricity to complex hearing and motion mechanisms – the observation and control of electromechanical responses on the nanoscale opens up a pathway for new advances in medicine, implants, bioMEMS, and other applications that we cannot yet envision," explained Nikiforov.
The researchers presented their work in Nanotechnology.