Russian politicians are increasingly using social networks publishing a lot of texts. One of the important issues in the context of the analysis of political online communication is the choice of negative and positive topics in publications as well as the reaction of the audience. In order to analyze the main patterns of this process we have collected the data from the social network Vkontakte. Our sample covers the period from 1 January, 2017 to 25 April, 2019, in total 46293 posts and 2197063 comments in 23 politician's accounts. To build the classifier we used two text corpora: Rubtsova's corpus and RuSentiment corpus. The algorithm of sentiment analysis was implemented on the basis of bidirectional recurrent neural network. Using Rubtsova's corpus we provided the accuracy of 91% and using RuSentiment we provided the accuracy of 84% (accuracy is calculated as the proportion of correctly identified cases from the test sample). We found that the markup of data significantly differs when different corpora were used. The most adequate results in the analysis of posts and comments, in our opinion, are obtained by using an ensemble of models based on the both corpora. As a result of classification, we identified a number of patterns. Thus, the number of likes and views of posts is higher for the posts classified as positive, and the number of reposts is higher for the posts classified as negative. We also found that the number of comments is higher for the posts with a negative sentiment, and the average sentiment of comments on positive posts is more positive than the average sentiment of comments on negative posts.