Prediction of protein stability changes upon one-point mutations using machine learning

This paper describes a new approach to the detection of protein stability change upon amino acid mutations. The main goal is to create a new meta-tool, which combines the outputs of eight well-established prediction tools and due to a suitable method of consensus making, it is able to improve the overall prediction accuracy. The optimal combination of outputs of these tools is found by using a various number of machine learning methods. Out of all tested machine learning methods, KStar showed the highest prediction accuracy on the training dataset compiled from experimentally validated mutations originating from ProTherm database. Due to this reason, it is chosen as an optimal consensus maker. The general prediction abilities are validated on the testing dataset composed of multi-point amino acid mutations extracted also from ProTherm database. Since the multi-point mutations were not used for training any of integrated tools, we suppose that such comparison is objective. As a result, the developed meta-tool based on KStar improves the correlation coefficient by 0.130 on the training dataset and 0.239 on the testing dataset, respectively (the comparison is being made against the most successful integrated tool). Based on the obtained results, we claim that machine learning methods may help identify strengths and weaknesses of contemporary protein stability prediction tools.