Cancer prognosis prediction and analysis can help patients understand expected life and help clinicians provide correct therapeutic guidance. Thanks to the development of sequencing technology, multi-omics data, and biological networks have been used for cancer prognosis prediction. Besides, graph neural networks can simultaneously consider multi-omics features and molecular interactions in biological networks, becoming mainstream in cancer prognosis prediction and analysis. However, the limited number of neighboring genes in biological networks restricts the accuracy of graph neural networks. To solve this problem, a local augmented graph convolutional network named LAGProg is proposed in this paper for cancer prognosis prediction and analysis. The process follows: first, given a patient's multi-omics data features and biological network, the corresponding augmented conditional variational autoencoder generates features. Then, the generated augmented features and the original features are fed into a cancer prognosis prediction model to complete the cancer prognosis prediction task. The conditional variational autoencoder consists of two parts: encoder-decoder. In the encoding phase, an encoder learns the conditional distribution of the multi-omics data. As a generative model, a decoder takes the conditional distribution and the original feature as inputs to generate the enhanced features. The cancer prognosis prediction model consists of a two-layer graph convolutional neural network and a Cox proportional risk network. The Cox proportional risk network consists of fully connected layers. Extensive experiments on 15 real-world datasets from TCGA demonstrated the effectiveness and efficiency of the proposed method in predicting cancer prognosis. LAGProg improved the C-index values by an average of 8.5% over the state-of-the-art graph neural network method. Moreover, we confirmed that the local augmentation technique could enhance the model's ability to represent multi-omics features, improve the model's robustness to missing multi-omics features, and prevent the model's over-smoothing during training. Finally, based on genes identified through differential expression analysis, we discovered 13 prognostic markers highly associated with breast cancer, among which ten genes have been proved by literature review.

Transcription factor binding sites (TFBSs) prediction is crucial for decoding cis-regulation. However, current deep learning methods fail to simultaneously consider the multiscale features from DNA sequences and histone modifications in an efficient manner. To this end, we propose a novel D ensely C onvolutional N eural N etwork using DNA S equence and H istone Modification, dubbed as DCNN-SH, for TFBSs prediction. Our model adopts densely convolutional blocks to reuse multi-length motifs and multi-order dependencies of nucleotides. This unique design allows our model to consider the multi-scale features using smaller convolutional kernels compared to current methods. Our work is the first to apply densely network for TFBSs prediction. Extensive experiments over 300 ChIP-seq datasets demonstrate that our model significantly outperforms several state-of-the-art prediction methods in terms of accuracy, ROC-AUC and PR-AUC.

To comprehend DNA-protein binding specificity in diverse cell types is essential to reveal regulatory mechanisms in biological processes. Recently, deep learning has been successfully applied to predict DNA-protein binding sites from large-scale chromatin-profiling data. However, the precise identification of putative binding sites in specific cell types with low labeled samples remains challenging. To this end, we present a novel Transferable Transformer-based method, dubbed as BindTransNet, for cross-cell types DNA-protein binding prediction. Transfer learning and Transformer Encoder are simultaneously adopted in our presented approach to capture some shared long-range dependencies between various motifs available in cross-cell types. This unique design helps our method recognize putative binding sites without massive labeled samples by leveraging the above-mentioned standard features. This work is the first to apply a Transformer for DNA-protein binding sites prediction. The presented method is measured on TFs COREST and SRF in four cell types with eight cell-type TF pairs. For both 4-class prediction and binary-level prediction, BindTransNet can significantly outperform several state-of-the-art methods. Moreover, BindTransNet achieves considerable margin performance improvements by leveraging transfer learning. This is a presuasive indication that BindTransNet can indeed capture shared features available in other cell types.