Transmembrane helix prediction using amino acid property features and latent semantic analysis

Background: Prediction of transmembrane (TM) helices by statistical methods suffers from lack of sufficient training data. Current best methods use hundreds or even thousands of free parameters in their models which are tuned to fit the little data available for training. Further, they are often restricted to the generally accepted topology "cytoplasmic-transmembrane-extracellular" and cannot adapt to membrane proteins that do not conform to this topology. Recent crystal structures of channel proteins have revealed novel architectures showing that the above topology may not be as universal as previously believed. Thus, there is a need for methods that can better predict TM helices even in novel topologies and families. Results: Here, we describe a new method "TMpro" to predict TM helices with high accuracy. To avoid overfitting to existing topologies, we have collapsed cytoplasmic and extracellular labels to a single state, non-TM. TMpro is a binary classifier which predicts TM or non-TM using multiple amino acid properties (charge, polarity, aromaticity, size and electronic properties) as features. The features are extracted from sequence information by applying the framework used for latent semantic analysis of text documents and are input to neural networks that learn the distinction between TM and non-TM segments. The model uses only 25 free parameters. In benchmark analysis TMpro achieves 95% segment F-score corresponding to 50% reduction in error rate compared to the best methods not requiring an evolutionary profile of a protein to be known. Performance is also improved when applied to more recent and larger high resolution datasets PDBTM and MPtopo. TMpro predictions in membrane proteins with unusual or disputed TM structure (K+ channel, aquaporin and HIV envelope glycoprotein) are discussed. Conclusion: TMpro uses very few free parameters in modeling TM segments as opposed to the very large number of free parameters used in state-of-the-art membrane prediction methods, yet achieves very high segment accuracies. This is highly advantageous considering that high resolution transmembrane information is available only for very few proteins. The greatest impact of TMpro is therefore expected in the prediction of TM segments in proteins with novel topologies. Further, the paper introduces a novel method of extracting features from protein sequence, namely that of latent semantic analysis model. The success of this approach in the current context suggests that it can find potential applications in other sequence-based analysis problems.