Publicação
Building an automated platform for the classification of peptides/proteins using machine learning
| Resumo: | One of the challenging problems in bioinformatics is to computationally characterize sequences, structures and functions of proteins. Sequence-derived structural and physico-chemical properties of proteins have been used in the development of machine learning models in protein related problems. However, tools and platforms to calculate features and perform Machine learning (ML) with proteins are scarce and have their limitations in terms of effectiveness, user-friendliness and capacity. Here, a generic modular automated platform for the classification of proteins based on their physicochemical properties using different ML algorithms is proposed. The tool developed, as a Python package, facilitates the major tasks of ML and includes modules to read and alter sequences, calculate protein features, preprocess datasets, execute feature reduction and selection, perform clustering, train and optimize ML models and make predictions. As it is modular, the user retains the power to alter the code to fit specific needs. This platform was tested to predict membrane active anticancer and antimicrobial peptides and further used to explore viral fusion peptides. Membrane-interacting peptides play a crucial role in several biological processes. Fusion peptides are a subclass found in enveloped viruses, that are particularly relevant for membrane fusion. Determining what are the properties that characterize fusion peptides and distinguishing them from other proteins is a very relevant scientific question with important technological implications. Using three different datasets composed by well annotated sequences, different feature extraction techniques and feature selection methods (resulting in a total of over 20 datasets), seven ML models were trained and tested, using cross validation for error estimation and grid search for model selection. The different models, feature sets and feature selection techniques were compared. The best models obtained for distinct metric were then used to predict the location of a known fusion peptide in a protein sequence from the Dengue virus. Feature importances were also analysed. The models obtained will be useful in future research, also providing a biological insight of the distinctive physicochemical characteristics of fusion peptides. This work presents a freely available tool to perform ML-based protein classification and the first global analysis and prediction of viral fusion peptides using ML, reinforcing the usability and importance of ML in protein classification problems. |
|---|---|
| Autores principais: | Sequeira, Ana Marta Fernandes Tavares |
| Assunto: | Aprendizagem máquina Classificação de péptidos Péptidos de fusão viral Machine Learning Peptide Classification Viral Fusion Peptides |
| Ano: | 2019 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | One of the challenging problems in bioinformatics is to computationally characterize sequences, structures and functions of proteins. Sequence-derived structural and physico-chemical properties of proteins have been used in the development of machine learning models in protein related problems. However, tools and platforms to calculate features and perform Machine learning (ML) with proteins are scarce and have their limitations in terms of effectiveness, user-friendliness and capacity. Here, a generic modular automated platform for the classification of proteins based on their physicochemical properties using different ML algorithms is proposed. The tool developed, as a Python package, facilitates the major tasks of ML and includes modules to read and alter sequences, calculate protein features, preprocess datasets, execute feature reduction and selection, perform clustering, train and optimize ML models and make predictions. As it is modular, the user retains the power to alter the code to fit specific needs. This platform was tested to predict membrane active anticancer and antimicrobial peptides and further used to explore viral fusion peptides. Membrane-interacting peptides play a crucial role in several biological processes. Fusion peptides are a subclass found in enveloped viruses, that are particularly relevant for membrane fusion. Determining what are the properties that characterize fusion peptides and distinguishing them from other proteins is a very relevant scientific question with important technological implications. Using three different datasets composed by well annotated sequences, different feature extraction techniques and feature selection methods (resulting in a total of over 20 datasets), seven ML models were trained and tested, using cross validation for error estimation and grid search for model selection. The different models, feature sets and feature selection techniques were compared. The best models obtained for distinct metric were then used to predict the location of a known fusion peptide in a protein sequence from the Dengue virus. Feature importances were also analysed. The models obtained will be useful in future research, also providing a biological insight of the distinctive physicochemical characteristics of fusion peptides. This work presents a freely available tool to perform ML-based protein classification and the first global analysis and prediction of viral fusion peptides using ML, reinforcing the usability and importance of ML in protein classification problems. |
|---|