Most of the sequence-based structure prediction and experimental structure determination research have been focused on soluble proteins, with relatively less attention paid to the transmembrane proteins. With the availability of high-quality methods for the structure prediction of the soluble proteins, there is a need to fill the gaps for the prediction of transmembrane proteins, motivated by the fact that they are the primary targets of many drug design efforts. We provide a systematic survey of machine learning methods that predict various aspects of topological properties of transmembrane proteins, including helical and beta-barrel proteins. We reveal that despite a smaller number of unique solved structures for this group of proteins, a battery of corresponding algorithms has been developed. These algorithms offer various degrees of accuracy which needs further improvements. We present a case study to illustrate a typical predictive scenario that shows that the available prediction methods may produce somewhat different results and we discuss how to deal with situations like this.