This paper is devoted to predicting the transmembrane helices in proteins by statistical modeling. A novel segment-training algorithm for Hidden Markov modeling based on the biological characters of transmembrane proteins has been introduced into training and predicting the topological characters of transmembrane helices such as location and orientation. Compared to the standard Balm-Welch training algorithm, this algorithm has lower complexity while prediction performance is better than or at least comparable to other existing methods. With a 10-fold cross-validation test on a database containing 160 transmembrane proteins, an HMM model trained with this algorithm outperformed two other prediction methods: TMHMM and MEMSTAT; the novel method was validated by its prediction sensitivity (97.0%) and correct location (91.3%). The results showed that this algorithm is an efficient and a reasonable supplement to modeling and prediction of transmembrane helices.