Transmembrane Proteins Based On DNN And CHMM Research On The Combination Prediction Of Topology And Secondary Structure

Transmembrane proteins are involved in many important physiological processes and are of great importance in drug design and disease treatment.Transmembrane proteins span lipid bilayers in alpha-helix or beta-barrel structures,and there are certain challenges in using biological experiments to determine their structure,resulting in a small number of transmembrane proteins with known structures.Due to the extensive biological significance of transmembrane proteins,it is important to predict their structures using computer biology methods.The topological structure of transmembrane proteins can determine the spatial conformation relative to the membrane,while the secondary structure helps to identify the functional domain.The secondary structure of transmembrane proteins has a high degree of consistency in the transmembrane region,and achieving combined structure prediction is a guide to further understanding the structure and function of transmembrane proteins.The studies have shown that the current popular protein structure predictor,Alphafold2,performs well in its structural prediction even without considering the specificity of transmembrane proteins.However,our combined prediction of transmembrane protein topology and secondary structure still needs to be explored.Due to the fact that structure determines function,these comprehensive prediction methods with strong performance help some work directly understand the potential regions of expression function from sequences,and providing intuitive references for related field research.Even though there have been more machine learning and deep learning methods to achieve transmembrane protein structure prediction,there is still room for improvement in their comprehensive performance and biological significance.The combination of machine and deep learning is worth exploring,including the combination of Long and Short-Term Memory networks with Conditional gravitational fields,combination of Neural Networks and Hidden Markov Model.In this paper,we combine Deep Learning Neural Networks(DNN)and Class Hidden Markov Model(CHMM)to construct hybrid models to achieve combined topological and secondary structure prediction of alpha-helix transmembrane proteins based on sequences,not only learning implicit features from biological sequences to identify complex patterns,but also using statistical models to capture state-associated features to construct models with high synthesis capability.In this paper,we propose separate and joint training methods to construct two prediction models including HDNNtopss and Co-HDNNtopss.The deep learning part uses HHblits and One-Hot feature encoding to extract features by multi-scale Convolutional Neural Networks;Bi-directional Long and Short-term Memory networks with attention enhancement is used to capture context dependence to obtain output probabilities.A CHMM replaces the firing probabilities with the output probabilities of the deep learning part to achieve a combination of the two types of methods.In this paper,the two models are analyzed in terms of merged structure,topology and secondary structure prediction performance against other corresponding state-of-the-art methods on an independent test set using strictly uniform evaluation metrics.The results show that the separate training method HDNNtopss in this paper has a strong performance in predicting the merged structure and topology of alpha-helix transmembrane proteins,and the secondary structure also reaches the general prediction level,and has a strong comprehensive performance in general.In contrast,the joint training method Co-HDNNtopss in this paper is not the best performance casually,but it provides the idea of joint training of deep learning and a Hidden Markov Model,which is informative for the training of such hybrid models.We conclude through analysis and case studies that there is still room for innovation and improvement in our structure prediction accuracy due to the imbalance of sample structure categories.