AntiMicrobial Resistance Prediction Using Advanced Image Encoding Techniques

Antimicrobial resistance (AMR) is one of the significant problems in the health sector and requires resources to create and develop accurate forecasting systems capable of recognizing resistant microbes and their treatment approaches. The traditional machine learning models used in AMR prediction usually suffer from poor genomic sequences feature extraction and finding it hard to solve big biological problems with big data. K-mer based methods or other simpler classifiers have been able to provide certain accuracy, but these have strived when dealing with many genomic strains and variants. To address these shortcomings, deep learning models are able to learn data representation at various abstraction levels which understanding complex data of fundamental and high dimensional patterns. In this work we provide a new deep learning approach exploiting CNNs on FCGR images of microbial genomes. The sequence is represented as frequency chaos game representation making it better for the enn structural application as resnet and densenet where features are deeply embedded in classifiers. When it comes to string and image based methods it is clear that they both have their own benefits and shortcomings. String based analysis is a faster and more accurate way to analyze naked sequence data conversion, but it may not be as informative as visual presentation. Our study verified the hypothesis that it is feasible to encode genomic data for further FCGR, so that the data might be utilized as an input for machine learning and deep learning methods. The model is effective with substantial improvements with the accuracy of over $90 \%$ in test dataset. This represents a significant improvement in accuracy score with better systems for predicting antibiotic resistance among many types of microbes.