pybabs

Machine Learning algorithms already support our daily life in many areas of application. Be it self-driving cars, intelligent robots, or the field of computational biology. Identi- fying specific features from biological sequences is essential in understanding the world surrounding us and its complex connections. Forensics, the development of new drugs, and tools for predicting certain Illnesses based on our DNA are important subareas of this Topic. One standard method to do so is in the form of Support Vector Machines (SVM). SVMs are well suited for this task for several reasons, but the biggest profit comes from their ability to use the so-called kernels for their calculation. They convert the alphabetical nature of biological sequences via similarity analysis into their own numerical similarity Matrices, called Kernel. This Kernel can then be used for internal calculations. In the case of biological sequences, this special feature clearly distinguishes the support vector machines from other machine learning algorithms This thesis is based on the KEBABS Package from Johannes Palme, which was im- plemented in R. Hence, the user base of R is declining, and the functionalities of the Kebabs Package are more and more sought after in other Programming languages. Hence Python is the go-to language for processing biological Data, it makes sense to provide an easy-to-use framework for manipulating and processing biological Data in this well-supported platform. The Package offers two Kernels for Kernel-Based Sequence analysis: the spectrum kernel and the gappy pair Kernel. Both can be used with DNA, RNA, and Aminoacids sequences. The Package also explicitly represents the calculated Kernel Matrix as a whole or in the sparse format. The Package also works seamlessly with multiple existing SVM Frameworks in the Python landscape, like Libsvm and the prevalent scikit learning library. The Package also provides Cross-validation, grid search, and an unbiased model se- lection. For better biological interpretability, the weights used for the calculations can easily be extracted. Also, prediction profiles have been implemented to understand bet- ter how some parts of a sequence contribute to the overall result.