Enhancer-Promoter Interactions (EPIs) play a crucial role in gene expression regulation. However, traditional experimental methods for detecting EPIs are often time-consuming and costly, prompting a growing demand for computational approaches. In this work, we propose a novel deep-learning model, termed EPI-RMDL, for the prediction of enhancer-promoter interactions based only on the DNA sequences. EPI-RMDL at first encodes the DNA sequences of a set of enhancers and promoters into an information matrix via the dna2vec method. Subsequently, local and global features are extracted from the matrix using a three-layer convolution neural network. Then, the features are processed through three RoFormers, which are enhanced transformers with Rotary Position Embedding (RoPE), in order to obtain relative positional information and interaction details between promoters and enhancers. Finally, a special matching mechanism is incorporated to analyze the interplay among the output vectors generated by the front-end RoFormers. We trained a general model by integrating data from six distinct cell lines and fine-tuned it with specific cell-line data to obtain an optimal model EPI-RMDL best for each cell line. Benchmarking against six state-of-the-art methods using datasets from six cell lines, our model demonstrates superior performance. Specifically, the EPI-RMDL_best model achieves a mean AUROC of 95.8% and an average AUPR of 80.8%.

Medication recommendation is a crucial application of artificial intelligence in healthcare. Current methodologies mostly depend on patient-level longitudinal representation, which utilizes the entirety of historical electronic health records for making predictions. However, they tend to overlook a few key elements: (1) The need to analyze the impact of past medications on previous conditions. (2) Similarity in patient visits is more common than similarity in the complete medical histories of patients. (3) It is difficult to accurately represent patient-level longitudinal data due to the varying numbers of visits. To our knowledge, current models face difficulties in dealing with initial patient visits (i.e. in cold-start scenarios) which are common in clinical practice. This paper introduces DrugDoctor, an innovative drug recommendation model crafted to emulate the decision-making mechanics of human doctors. Unlike previous methods, DrugDoctor explores the visit-level relationship between prescriptions and diseases while considering the impact of past prescriptions on the patient's condition to provide more accurate recommendations. We design a plug-and-play block to effectively capture drug substructure-aware disease information and effectiveness-aware medication information, employing cross-attention and multi-head self-attention mechanisms. Furthermore, DrugDoctor adopts a fundamentally new visit-level training strategy, aligning more closely with the practices of doctors. Extensive experiments conducted on the MIMIC-III and MIMIC-IV datasets demonstrate that DrugDoctor outperforms 10 other state-of-the-art methods in terms of Jaccard, F1-score, and PRAUC. Moreover, DrugDoctor exhibits strong robustness in handling patients with varying numbers of visits and effectively tackles "cold-start" issues in medication combination recommendations.

MOTIVATION: The identification of cancer subtypes plays a crucial role in cancer research and treatment. With the rapid development of high-throughput sequencing technologies, there has been an exponential accumulation of cancer multi-omics data. Integrating multi-omics data has emerged as a cost-effective and efficient strategy for cancer subtyping. While current methods primarily rely on genomics data, protein expression data offers a closer representation of phenotype. Therefore, integrating protein expression data holds promise for enhancing subtyping accuracy. However, the scarcity of protein expression data compared to genomics data presents a challenge in its direct incorporation into existing methods. Moreover, striking a balance between omics-specific learning and cross-omics learning remains a prevalent challenge in current multi-omics integration methods. RESULTS: We introduce Subtype-MGTP, a novel cancer subtyping framework based on the translation of Multiple Genomics To Proteomics. Subtype-MGTP comprises two modules: a translation module, which leverages available protein data to translate multi-type genomics data into predicted protein expression data, and an improved deep subspace clustering module, which integrates contrastive learning to cluster the predicted protein data, yielding refined subtyping results. Extensive experiments conducted on benchmark datasets demonstrate that Subtype-MGTP outperforms nine state-of-the-art cancer subtyping methods. The interpretability of clustering results is further supported by the clinical and survival analysis. Subtype-MGTP also exhibits strong robustness against varying rates of missing protein data and demonstrates distinct advantages in integrating multi-omics data with imbalanced multi-omics data. AVAILABILITY AND IMPLEMENTATION: The code and results are available at https://github.com/kybinn/Subtype-MGTP.

Identifying cancer subtypes is an essential component of precision medicine, as it helps researchers develop more precise treatment methods and prevention strategies. Meanwhile, high-throughput sequencing technologies have produced a huge amount of multi-omics data for cancer patients and make it is practical to subtype cancers using multi-omics data. As existing cancer subtyping computational models based on multi-omics data could not effectively extended to weakly paired omics data, we proposed a novel unsupervised cancer subtyping model Subtype-DCGCN. Subtype-DCGCN uses Dual Contrast Graph Convolutional Networks guided by dual contrastive learning to lean low dimensional features for each type omics data, and with weighted average fusion Subtype-DCGCN could deal well with weakly paired multi-omics data. Extensive experiments on benchmark datasets showed that Subtype-DCGCN exhibited superior performance to other eight state-of-the-art similar methods in general to identify cancer subtypes. Moreover, tests on simulated datasets with varying missing rate showed that Subtype-DCGCN performed pretty well on weakly paired omics datasets.