A Transcription Factor Binding Site Prediction Algorithm Based on Semi-Supervised Learning

Transcription factor binding site (TFBS) is a DNA sequence that binds to transcription factors and interacts with transcription factors to regulate the transcription process of genes. Most prediction methods for TFBS use artificial synthesis or interception of unanalyzed DNA fragments as negative samples in the model, so the prediction results with negative samples are rich in noise and will affect the performance of the classification model. Therefore, this paper proposes a model based on deeply convolutional autoencoders using only experimentally validated transcription factor binding site dataset. The prediction model uses a convolutional neural network with multiple hidden layers and classification layer, and the classification layer uses the sigmoid classifier; In addition, unsupervised learning and supervised learning are adopted as training strategies for predictive models. The model uses the ChIP-Seq data processed by the Gene Transcription Regulation Database (GTRD) as the experimental data set. The human and mouse datasets as independent testing, AUC and ACC were as high as 86% and 94%, respectively, the experiments have shown that a single model is superior to specific models in training 9 human transcription factors and 6 mouse transcription factors. The experimental results show that the proposed method is very feasible and effective in the prediction of TFBS.