Transformer-based models have achieved stateof-the-art performance in document classification but struggle with long-text processing due to the quadratic computational complexity in the self-attention module.Existing solutions, such as sparse attention, hierarchical models, and key sentence extraction, partially address the issue but still fall short when the input sequence is exceptionally lengthy.To address this challenge, we propose IRIS (Interpretable Retrieval-Augmented Classification for long Interspersed Document Sequences), a novel, lightweight framework that utilizes retrieval to efficiently classify long documents while enhancing interpretability.IRIS segments documents into chunks, stores their embeddings in a vector database, and retrieves those most relevant to a given task using learnable query vectors.A linear attention mechanism then aggregates the retrieved embeddings for classification, allowing the model to process arbitrarily long documents without increasing computational cost and remaining trainable on a single GPU.Our experiments across six datasets show that IRIS achieves comparable performance to baseline models on standard benchmarks, and excels in three clinical note disease risk prediction tasks where documents are extremely long and key information is sparse.Furthermore, IRIS provides global interpretability by revealing a clear summary of key risk factors identified by the model.These findings highlight the potential of IRIS as an efficient and interpretable solution for long-document classification, particularly in healthcare applications where both performance and explainability are crucial.1

Rapid advances in genomics discovery tools and a growing realization of microRNA's implication in intercellular communication have led to a proliferation of studies of circulating microRNA sorting and regulation across cells and different species. Although sometimes, reaching controversial scientific discoveries and conclusions, these studies have yielded new insights in the functional roles of circulating microRNA and a plethora of analytical methods and tools. Here, we consider this body of work in light of key computational principles underpinning discovery of circulating microRNAs in terms of their sorting and targeting, with the goal of providing practical guidance for applications that is focused on the design and analysis of circulating microRNAs and their context-dependent regulation. We survey a broad range of informatics methods and tools that are available to the researcher, discuss their key features, applications and various unsolved problems and close this review with prospects and broader implication of this field.

MicroRNAs regulate virtually the whole gene network in human body and have been implicated in most physiological and pathological conditions including cancers. Understanding the precise mechanisms of microRNA-mRNA interaction is fundamentally important to elucidate the important roles of miRNA in regulating various cellular and disease developmental stages. Numerous computational methods have been developed for miRNA target prediction, mostly focusing on static binding prediction and highly dependent on sequence-pairing interactions. However, the interplay between competing and cooperative microRNA-target binding makes it exceptionally complex and challenging for reliable target identification, which has hindered the existing tools from practical use. In this study, we present a new computational method for microRNA target prediction using the Dirichlet Process Gaussian Mixture Model (DPGMM). A comprehensive collection of features related to sequence and structure of microRNAs, mRNAs, and the binding sites have been assessed to optimize the statistical prediction of new binding sites in human transcriptome. Through multiple evaluations on recently-discovered miRNA-mRNA interactions reported in large-scale sequencing analyses and a screening test on the entire human transcripts, the results show that our model outperformed several state-of-the-art tools in terms of reduced false positive prediction and promising predictive power on binding sites specific to transcript isoforms.

MicroRNA is a type of short non-coding RNAs, which post-transcriptionally regulate gene expressions. It has been well-documented that human microRNAs contribute in the disease development, such as cancers and obesity. While most microRNA functional studies heavily rely on the regulatory interactions between microRNAs and their target messenger RNAs, the accumulating evidence has shown that the altered availability of microRNAs, their target, and other types of endogenous RNAs competing are able to affect the microRNA-target interactions efficiently, which reflects the dynamic and conditional property of microRNA-mediated gene regulation. Here we present a new computational pipeline, miRDR-OSG, that utilizing the high-throughput computing resource provided by Open Science Grid to study the dynamic regulation of microRNAs in cancer development with the consideration of both competing and cooperative mechanisms. A large-scale genomic dataset from over four thousand patients with 9 major types of cancer was used to demonstrate the usage of miRDR-OSG. As a result, we identified 10,726 microRNA regulatory interactions that only occurred in a specific stage and/or cancer type. This observation demonstrated the dynamic and conditional microRNA regulation during cancer progression. miRDR-OSG is freely available at http://sbbi.unl.edu/miRDR.

MicroRNAs are functionally important non-coding RNAs that silence host genes via destabilizing the mRNAs or preventing the translation. Given the far-reaching implication of microRNA regulation in human health, novel bioinformatics tools are desired to facilitate the mechanistic understanding of microRNA mediated gene regulation and its regulatory roles in human disease development. Most state-of-the-art computational methods focus on the functional inference through identifying microRNA targets and none has comprehensively investigated the functional similarity among microRNAs. Here we present a new method to quantitatively measure the functional relevance among microRNAs that is solely based on the Gene Ontology and integrated functional annotation data from public pathways and PFam gene databases. In this study, we further demonstrated the use of the derived microRNA pairwise similarities to investigate the cooperative microRNA modules and construct the genome scale microRNA mediated gene network in human. The complete results and the similarity assessment system can be freely accessed at (http://sbbi.unl.edu/microRNASim).

MicroRNAs, a class of short non-coding RNAs, are able to regulate more than half of human genes and affect many fundamental biological processes. It has been long considered synthesized endogenously until very recent discoveries showing that human can absorb exogenous microRNAs from dietary resources. This finding has raised a challenge scientific question: which exogenous microRNAs can be integrated into human circulation and possibly exert functions in human? Here we present a well-designed ensemble manifold ranking model for identifying human absorbable exogenous miRNAs from 14 common dietary species. Specifically, we have analyzed 4,910 dietary microRNAs with 1,120 features derived based on the microRNA sequence and structure. In total, 70 discriminative features were selected to characterize the circulating microRNAs in human and have been used to infer the possibility of a certain exogenous microRNA getting integrated into human circulation. Finally, 461 dietary microRNAs have been identified as transportable exogenous microRNAs. To assess the performance of our ensemble model, we have validated the top predictions through a milk-feeding study. In addition, 26 microRNAs from two virus species were predicted as transportable and have been validated in two external experiments. The results demonstrate the data-driven computational model is highly promising to study transportable microRNAs while bypassing the complex mechanistic details.