Abstract:
Nanopore detection technology plays important role in the field of single molecule detection. The feature extraction and transformation of nanopore blockade current is the key to classify signals and determine the analytes. Since the supervised learning can only predict the blockade current with known types, the unsupervised learning would be expected to achieve the distinction of the essential characteristics of signals. By means of the blockade current based on DNA carrier and the feature extraction of convolutional neural network, this paper proposes an unsupervised clustering scheme for nanopore signals. The proposed method combines the improved deep embedded clustering with the variational auto-encoder so that the feature transformation and the clustering of the blockade current can be jointly trained. In the original improved deep embedded clustering, the model utilizes an auto-encoder to transform the input data to a new compressed representation, which can reserve the crucial information of the input data. Thus, the auto-encoder may be replaced by a variational auto-encoder to learn the parameters of a probability distribution on the input data, instead of just learning the compressed representation. Finally, the proposed clustering model is firstly verified via 8 clustering centroids. Due to the symmetrical encoded DNA, like “100” and “001”, the case of 6 clustering centroids is further conducted to validate the proposed clustering model. It is shown from the experiment results that the proposed method can provide better clustering results for blockade current of DNA carrier, and achieves a 29% boost at most, compared with other clustering algorithms. .