Monday 16:10–16:40 in Track 3

Using convolutional neural networks to analyze bacteriophages DNA

Michał Jadczuk

Audience level:
Intermediate

Description

With the overuse of antibiotics causing the emergence of superbugs, people all over the world are looking into their safer alternatives. Bacteriophages - or viruses that feed on the bacteria - might just be one of them. All phage strains must be carefully analyzed before commercial use, though. This presentation will explain how machine learning can speed up the evaluation of bacteriophages DNA.

Abstract

The bacteriophages are viruses feed on bacteria that are viewed as a potential alternative to the antibiotics. One could say that antibiotics work like “bombs” – they kill both good and bad bacteria in your system, which isn’t ideal. On the other hand, the so-called phage cocktails can target only the specific strains of bacteria that cause the actual treated disease, which potentially makes them much safer. Also, the overuse of antibiotics in the agriculture industry causes the emergence of the “superbugs”, which are drug-resistant bacteria – not to mention that antibiotics can cause other side effects upon meat consumption. Commercialized bacteriophage-based alternatives could potentially prevent that.

However, before commercializing a certain bacteriophage strain, one has to perform a series of tests to make sure it meets all the necessary criteria. The tests might take weeks or even months of tedious laboratory work. One of the most important criteria that the phages must meet, right next to the bacteria type they target, is the way they can reproduce. Based on the reproduction cycle, one can detect two distinct types of phages: lytic and lysogenic. Only the lytic bacteriophages can be used in phage cocktails, as lysogenic strains might cause bacteria mutations.

Detecting whether the strain is lytic or lysogenic used to require manual work, but we managed to speed up this process. All thanks to developing an automated classifier that can detect the reproduction cycle type using bacteriophage genome as input processed by the convolutional neural networks. This lecture will describe the origin of the project, difficulties with processing DNA, and – more importantly – it will try to prove that convolutional neural networks are not only effective for images, but can also handle sequential data (such as the DNA) reasonably well.

Subscribe to Receive PyData Updates

Subscribe