
Supervised machine learning (ML) is widely used for noncoding RNA (ncRNA) classification; however, model performance is highly dependent on the construction of negative training datasets. Common strategies, such as random or dinucleotide-preserving shuffling, may introduce biases that lead to overestimated performance and reduced generalizability. Despite this, there is currently no standardized and accessible framework for generating and comparing negative RNA datasets. We present NeRNA-Web, a user-friendly web-based platform designed for generating structure-aware negative RNA sequences and conducting comparative evaluations. NeRNA-Web implements the previously developed NeRNA transformation, which incorporates predicted secondary-structure information into an encoded representation to generate transformed negative sequences. In addition to NeRNA-based generation, the platform incorporates several commonly used negative generation methods, such as random shuffling and dinucleotide-preserving shuffling, all within a unified and reproducible pipeline. The platform supports various ncRNA classes, including miRNA, lncRNA, tRNA, and circRNA, and offers visualization and export functionalities for downstream analyses. NeRNA-Web provides an accessible and extensible framework for generating and comparatively characterizing negative RNA datasets using alternative generation strategies. By enabling users to examine and download the generated datasets, the platform facilitates transparent and reproducible negative-data construction for downstream ncRNA research. The observed differences among generation strategies further emphasize the importance of carefully selecting and evaluating negative datasets according to the requirements of the intended downstream analysis.
Precise prediction of DTIs (drug-target interactions) is essential for drug repositioning and novel therapeutic development. Despite progress in deep learning and computing power, DTI prediction remains a time-consuming task. And most existing approaches overlook the influence of differentially weighted residual attention mechanisms on model performance, which could hinder their predictive performance. To address these limitations, we present DWRA-DTI (Deep Weighted Residual Attention-DTI), a novel DTI prediction framework that integrates streamlined preprocessing and early stopping strategies to achieve sub-minute inference times without sacrificing accuracy.Specifically, drug and protein feature matrices are encoded into embeddings of a unified dimension using graph convolutional encoders, which are subsequently fused through a weighted residual attention module to capture intricate interaction patterns. The fused representations are then passed into a MLP (Multilayer Perceptron) to predict potential interactions. Our study investigates the impact of varying weights in residual attention through ablation experiments, demonstrating that weighted residual attention effectively enhances model performance. The proposed model achieves outstanding results across multiple datasets, offering a promising approach and valuable insights for advancing current drug-target interaction prediction methods.
Exploring the fungal circular RNA (circRNA) landscape is bottlenecked by experimental and computational limits. Standard mRNA-seq systematically discards circRNAs due to their lack of poly(A) tails, while total RNA-seq remains too costly for large-scale screening. Existing computational models trained exclusively on human or plant sequences fail in fungi due to extreme genomic divergence across lineages, which range from intron-poor yeasts to intron-rich filamentous species. Consequently, no dedicated framework currently exists for de novo fungal circRNA prediction, leaving most non-model fungi unexplored. We present mycoCirc, the first end-to-end pan-fungal multi-modal pretrained model for fungal circRNA prediction. It integrates three data modalities, namely genome sequence, gene annotation and junction-level sequence, plus a species identifier, and optionally host-gene expression during fine-tuning, through a hierarchical fusion architecture that includes bidirectional cross-attention to capture donor–acceptor site interactions. Pretrained on 17 training strains spanning 14 fungal species and three taxonomic groups, mycoCirc was fine-tuned per group using leave-one-strain-out cross-validation. On held-out test species under Mode A (Genome + GTF, no RNA-seq), mycoCirc achieved an AUROC of 0.695–0.738, substantially outperforming JEDI (0.51–0.57) and CircPCBL both as published with plant-pretrained weights (0.45–0.57) and when retrained from scratch on fungal data (0.52–0.61). Cross-species evaluation on four independent fungi datasets demonstrated moderate but consistent cross-species transfer (AUROC 0.59–0.70 across all model–species combinations). Furthermore, its JunctionEncoder module enables back-splicing junction ranking with group-dependent accuracy (top-1 accuracy of 90.8
Long-read sequencing enables direct detection of DNA methylation at single-molecule resolution, but comprehensive analysis remains challenging because it requires multiple processing steps. Existing workflows are often limited to a single sequencing platform or provide limited support for downstream methylation analysis. Here, we present nf-core/methylong, an end-to-end Nextflow pipeline for detecting and analyzing 5-methylcytosine (5mC) and N6-methyladenine (6mA) from Oxford Nanopore Technologies (ONT) and PacBio data. The pipeline integrates methylation calling, alignment, methylation pileup and Fiber-seq analysis, supporting haplotype-level and population-scale differential methylation analysis. Benchmarking against whole-genome bisulfite sequencing yielded site-level Pearson correlations of 0.880-0.976 across integrated callers. Haplotype-level DMRs overlapped known imprinted intervals, while Fiber-seq analysis revealed nucleosome-associated 6mA patterns in both ONT and PacBio data. We further developed methylongAgent, a companion agent that assists with parameter selection, command generation and result summarization. nf-core/methylong standardizes long-read DNA methylation analysis across sequencing platforms, while methylongAgent further reduces the manual effort required for workflow configuration, execution and result interpretation, supporting scalable and interpretable long-read methylation analysis.
Understanding protein desiccation tolerance is fundamental for deciphering how organisms survive extreme dehydration. While extensive computational tools exist for predicting aggregation, solubility and thermostability, they capture only part of the properties relevant to the dry state. Thus, a specialized computational approach is required to identify these distinct tolerance features. We present a high-throughput deep learning framework that predicts protein desiccation tolerance by leveraging the ESM-2 protein language model. By capturing high-dimensional sequence representations, our model achieves competitive predictive accuracy without the computational overhead of explicit structural modeling. On a held-out test set, our method outperformed conventional predictors for aggregation, solubility, and thermostability, indicating that desiccation tolerance reflects a specific combination of these properties. Independent experimental datasets indicate that the model reproduces the reported ranking of proteins under both desiccation and lyophilization. Further analysis reveals that this tolerance is associated with predicted structural order and an acidic surface bias, and that the model transfers to the proteomes of other species. Desiccation tolerance is a composite, sequence-encoded biophysical property that can be predicted directly from amino acid sequences. The framework enables proteome-wide screening of desiccation-tolerant candidates, offering a tool for engineering drought-resistant crops and for stabilizing biopharmaceuticals for cold-chain-free storage.