| The small molecular compounds that are the most basic substances that re alize the physiological functions of organisms are called metabolites.It i s inseparable from the physiological state and function of the life body,a nd can directly reflect the physiological and biochemical reactions carried out in the life body.It contains prints that analyze the types and amounts of metabolites during the life of a living body,revealing the biochemical mechanism behind it.Compared with traditional omics research,metabolomics can more fully show the true physiological state and processes of living bo dies.As mentioned above,its research is widely used in the screening of b iomarkers,toxicology research,drug design and environmental science.In in any fields such as research.The signal data obtained by the collection and detection of metabolites is called the metabolome feature data and is the basic object of metabolome research.In order to extract the characteristic biological information ins ide,we often use statistical analysis and shallow machine learning methods for processing.However,modern metabolomics data has the following charact eristics:1.High dimensions and relatively few samples;2.The data contai ns a lot of noise,and only a few features are highly correlated with the r esearch object,and traditional methods are often not satisfactory Analysis results.Therefore,researchers have introduced feature selection algorithm s to preprocess the input data.These methods essentially add a weight valu e to the variable.Studies have shown that metabolites and metabolites are not completely unrelated.We need to use a more intelligent feature selecti on method to take into account the connection between substances during fea ture selection.Especially at present,there are a large number of communit y cohort studies at home and abroad.These studies can generate a large amo unt of subclinical data.Compared with traditional clinical data,the signa l noise intensity and the amount of information are more serious..In orde r to come up with these data,we used deep learning methods to filter the f eature data.Metabolomic data has a high feature dimension,especially sub-clinical d ata is generally non-specific,and often requires a very comprehensive tes t,and sometimes even uses several detection modes,its feature extraction data is complex and large-scale optimization problems.In this paper,we u se two-layer migration convolutional neural network to deal with it effecti vely.Convolutional neural networks and transfer learning strategies are im portant components in the field of deep learning.By reasonably allocating the process of global optimization and the side rate of local search,the c onvolutional neural network can obtain better feature extraction results th an other algorithms in a smaller space and time complexity.After in-depth analysis of the current popular convolutional neural network learning metho ds,we propose to use the metabolome data of clinically diagnosed patients for training,and then transfer the training results to the subclinical fie 1d,and then conduct a deeper round of learning,thus Obtained better featu re extraction results than existing metabolomic feature data processing met hods.By introducing deep machine learning algorithms into the framework in a encapsulated form to evaluate training performance,the paper proposes an i ntelligent feature extraction algorithm for sub-clinical high-dimensional m etabolomics feature data.In feature data analysis,better prediction resul ts are obtained than traditional learning methods.The substance label extr acted by the special fruit feature extractor effectively explains the assoc iation between the metabolite and the target biological physiological stat e.In addition,the model trained through this data set has strong multiple xing performance,and can be directly transferred in related research in th e future.It can be used as a complete toolkit for metabolomics research fo r subsequent further research. |
PDF Full Text Request