Spatial omics studies compare cell-cell organization across samples, but most methods model between-sample variability while treating sample-level spatial estimates as error-free. Overlooking within-sample uncertainty can distort inference in heterogeneous cohorts, motivating methods that explicitly quantify and propagate this uncertainty into cohort-level analyses.

Modern sequencing technologies can now capture multiple omic layers from the same biological system, but integrating these views into a coherent model is far from trivial. Graph-based deep learning has become an attractive strategy because it can represent complex molecular interactions and sample relationships in a flexible way. In this review, we survey how graphs are constructed and used in multi-omics deep learning models, organizing methods by node schema, edge semantics, interaction type, integration strategy, graph context, and model architecture across bulk, single-cell, and spatial settings. We summarize the strengths and weaknesses of different design choices in terms of interpretability, data requirements, robustness to noise and missing modalities, and suitability for tasks ranging from prediction to mechanism-oriented discovery. Based on these insights, we outline a general, practical pipeline for constructing, curating, and evaluating graphs that can serve as a starting point for new multi-omics studies.

Protein function prediction is essential and fundamental for drug discovery and disease treatment. In recent years, deep learning methods have achieved notable improvements by exploiting either protein sequence or structural features. Specifically, Convolutional Neural Networks often fail to apprehend global protein topologies due to restricted receptive fields, while Graph Convolutional Networks excel at processing graph-structured data, a singular network paradigm fundamentally lacks the capacity to fully integrate diverse, multimodal features. Furthermore, combining modalities via static feature aggregation frequently limits model efficacy and causes modality interference. To resolve these challenges, we propose the Structure-Sequence Adaptive Synergy network (SSAS-GO), which employs a Multi-Scale Motif Block to extract localized sequence semantic anchors, alongside a parallel Dual-Stream Graph Encoder to capture spatial topologies. Subsequently, these representations are fed into a Task-Adaptive Cross-Modal gating mechanism. This core module dynamically recalibrates the weights of sequence and structural features. On the PDBch test set, SSAS-GO leverages native topologies to achieve state-of-the-art Area Under the Precision-Recall Curve (AUPR) scores of 0.463 for Biological Process and 0.559 for Cellular Component. Remarkably, on the AFch test set, the model achieves a substantial 17.7% relative AUPR improvement for Molecular Function tasks.

The European Union (EU) enforces strict regulations on the traceability and labeling of genetically modified organisms (GMOs), including genome-edited (GE) lines produced through new genomic techniques (NGTs). Identifying GE organisms created by single nucleotide variations (SNVs) is however challenging, as a single SNV alone cannot unambiguously define a GE line. Recently, we introduced the concept of generating a genetic fingerprint to distinguish a specific GE rice line. This proof-of-concept approach integrated whole-genome sequencing (WGS)-based characterization with the Illumina technology, the public 3 K Rice Genomes (3KRG) database, and statistical feature-selection tools, to select and combine key genetic elements, including GE on-target site(s) and cultivar-specific 2-SNV barcodes, into a unique genetic fingerprint. In the present study, we expand this concept into a generalized data-driven framework allowing identification of multiple rice lines. Supported by newly developed bioinformatics and statistical feature-selection-based pipelines, this optimized strategy enables the generation of genetic fingerprints irrespective of a rice cultivar's inclusion in publicly available databases like 3KRG. In addition, this refined strategy can leverage WGS data generated from both Illumina and Oxford Nanopore Technologies (ONT) platforms for fingerprint generation and GE line identification. Using two distinct in-house GE rice lines from different cultivars, along with various publicly available WGS datasets, we demonstrated the robustness, scalability, and specificity of this approach for reliable GE rice line identification. Our findings provide a methodological foundation for data-driven traceability of GE rice lines, reinforcing regulatory compliance, supporting intellectual property (IP) protection, and contributing to the responsible implementation of EU GMO/NGT legislation.

Complex traits and diseases are highly polygenic and understanding the full set of genes involved is a central challenge in biomedicine. However, due to sample size limitations and noise (technical and biological), experimental approaches for disease-gene discovery such as transcriptomics and GWAS result in long, noisy, heterogeneous gene lists, which may be trimmed to a subset of likely relevant genes while leaving several false negatives. Computational gene classification approaches, especially those using genome-scale molecular interaction networks, are promising avenues for complementing such experimental findings by analytically expanding observed gene lists based on the functional relatedness between genes. We previously introduced the network-based gene classification approach, GenePlexus, which was rigorously benchmarked to show state-of-the-art performance, especially for predicting novel genes associated with biological processes and fine-grained phenotypes. Network-based gene classification performance, however, declines for diseases, especially when the inputs are omics and GWAS-based long gene lists.

Boolean networks are a widely used modeling framework in systems biology for studying gene regulation, signal transduction, and cellular decision-making. Empirical studies indicate that biological Boolean networks exhibit a high degree of canalization, a property of Boolean update rules that stabilizes dynamics and constrains state transitions. Despite its central role, existing software packages provide limited support for the systematic generation of Boolean functions and networks with prescribed canalization properties.

Genomic and protein foundation models (GFMs and PFMs) have demonstrated strong performance in learning the language of DNA and proteins, but their use in large-scale sequence generation is limited by the latency of autoregressive decoding. Because every token triggers a forward pass of a large Transformer, whose inference is relatively slow, long-sequence generation quickly becomes costly.

Reference genomes are foundational to genomics but suffer from widespread ambiguity and incompatibility due to inconsistent naming, undocumented differences, and lack of formal mechanisms for comparison.

With recent advances in spatial transcriptomic technologies, multi-tissue section datasets are proliferating. While existing computational methods have achieved substantial progress in integrating multiple sections and correcting for batch effects, current approaches for spatial domain identification often fail to fully leverage both spatial context and gene expression information across consecutive sections. Moreover, prevailing graph contrastive learning frameworks typically depend on the construction of positive and negative sample pairs-a process susceptible to the introduction of noise. To overcome these limitations, we introduce STGAT, a framework that first achieves precise spatial alignment across sections using gene expression similarity. Within a unified spatial domain, STGAT employs a graph contrastive learning strategy that requires only positive pairs, enabling effective self-supervised representation learning of graph nodes. Experimental results demonstrate that STGAT effectively enhances clustering accuracy in spatial domain identification tasks across multi-section and cross-technology datasets. When applied to mouse olfactory bulb sections, the method yields sharply defined spatial domain boundaries and allows accurate identification of distinct anatomical regions. Furthermore, STGAT provides a more refined characterization of the tumor microenvironment. The source code used in this paper can be found in https://github.com/Jinsl-lab/STGAT.

Gene network enrichment analysis (GNEA) offers a robust approach for interpreting the complex molecular mechanisms underlying phenotypic variability. Despite its utility, prevailing GNEA methodologies are predominantly optimized for binary phenotypes, leading to substantial information loss when applied to continuous biological traits such as drug sensitivity, cancer progression, etc. Additionally, traditional enrichment strategies often exhibit conceptual inconsistency between their null models and test hypotheses, as they rely on permuting phenotype labels rather than genes. To address these limitations, we introduce cell line-specific gene network enrichment analysis (CellGNEA), a computational strategy designed to identify pathway-level molecular interactions associated with continuous phenotypes in a cell line-specific manner. CellGNEA constructs gene regulatory networks tailored to individual cell lines and assesses molecular interplays within these networks by integrating multiple network-derived metrics, including clustering coefficient, PageRank, and regulatory effects. Associations between gene networks and continuous phenotypes are quantified using a Kolmogorov-Smirnov-based statistic, with statistical significance determined via a gene permutation strategy that aligns the null model with the tested hypothesis. Monte Carlo simulation studies indicate that CellGNEA exhibits robustness and enhanced sensitivity in detecting network enrichment linked to continuous phenotypes. Applications of CellGNEA to drug sensitivity-specific gene networks enable the identification of leukemia-related pathways with molecular interactions significantly associated with therapeutic response. Notably, our analysis revealed that imatinib, quizartinib, and ruxolitinib consistently correlate with network-level remodeling across acute myeloid leukemia, myelodysplastic syndrome, and chronic myeloid leukemia pathways, and identified resistance-associated genes including PTPN11, MS4A1, and BTK. Overall, CellGNEA establishes a systematic, scalable framework for functional network analysis of continuous phenotypes, facilitating comprehensive characterization of cell line-specific biological properties and providing valuable insights for systems biology and precision medicine.

Single-cell RNA sequencing (scRNA-seq) enables genome-wide gene expression profiling at single-cell resolution but loses the spatial context essential for interpreting cell identity and tissue organization. In contrast, spatially resolved transcriptomics (SRT) preserves spatial information but typically lacks single-cell resolution or complete transcriptome coverage. To obtain a more comprehensive view of heterogeneous spatial domains and cellular gene expression, we present Cell2Map, an unsupervised deep learning method that integrates scRNA-seq and SRT data from the same tissue region. Cell2Map assigns individual cells to SRT spots using a graph attention autoencoder equipped with a specially designed multi-term objective function that jointly optimizes expression-based, density-based, and embedding-level similarity and distance constraints. On benchmark datasets from mouse cerebellum and hippocampus, Cell2Map achieves higher single-cell mapping precision and overall accuracy than three popular methods (Celloc, CytoSPACE, Tangram) across a range of noise levels and spot cell densities. In real cancer applications, Cell2Map resolves intratumoral heterogeneity by accurately localizing tumor subclones and separating normal epithelial cells from ductal carcinoma in situ regions, and more faithfully reconstructs tumor microenvironments and immune-cell localization than competing approaches. Across breast cancer and myocardial infarction datasets, Cell2Map consistently attains higher sensitivity with fewer false positives, in close agreement with histological and biological annotations.

Accurate reconstruction and quantification of full-length RNA isoforms remain challenging in long-read RNA sequencing due to sequencing artifacts, complex splicing, and incomplete annotations. Although Pacific Biosciences circular consensus sequencing (PacBio CCS) provides high-fidelity long reads, scalable and annotation-flexible analysis frameworks remain limited. Here, we present ISAtools, an efficient framework specifically designed for PacBio CCS data. ISAtools introduces a splice site chain representation that unifies read alignments and transcript annotations into a compact format for scalable isoform reconstruction. It further integrates unsupervised density-based clustering for transcription start and end site detection and a truncation-aware quantification strategy. Benchmarking on simulated, SIRV spike-in controls, and biological datasets shows that ISAtools accurately reconstructs both annotated and novel isoforms with reliable splice structures and transcript boundaries while maintaining high computational efficiency. ISAtools processes up to 80 million reads in ~21 min using ~6 GB memory and scales to nearly 1 billion reads in under 4 h with ~8 GB memory, supporting large-scale PacBio Iso-Seq studies.

The accurate deconvolution of bulk transcriptomes typically confounds stable physical cell identities with dynamic physiological states, limiting our understanding of complex microenvironmental processes such as ovarian aging. To overcome this, we present DeepMCD, an end-to-end multi-task deep learning framework designed to simultaneously deconvolve cell-type proportions and programmed cell death (PCD) compositional fractions from standard bulk RNA-seq data. By mapping high-dimensional expression profiles into a shared, tokenized latent space, DeepMCD employs a Transformer-based cross-task attention mechanism to explicitly leverage cellular morphological context for calibrating PCD predictions. Concurrently, an adaptive uncertainty-weighting loss ensures balanced optimization, effectively mitigating negative transfer. Extensive benchmarking demonstrates that DeepMCD significantly outperforms state-of-the-art single-task algorithms. Through rigorous ablation and interpretability analyses, we computationally substantiate the biological premise that functional death states are heavily reliant on specific cell-type contexts. Applying DeepMCD to real-world mouse ovarian aging cohorts, we reconstructed a cell-type-specific PCD landscape, bypassing the need for costly single-cell sequencing. Specifically, we identified an age-associated increase in inflammatory and lytic death modalities, together with a strong association between macrophage enrichment and pyroptosis during ovarian aging. Crucially, the DeepMCD-derived PCD fractions exhibit profound divergent correlations with core ovarian fibrosis-related genes. These computationally extracted signatures may serve as cost-effective candidate digital biomarkers for evaluating ovarian fibrosis and reproductive senescence. Ultimately, DeepMCD provides a highly interpretable, robust, and scalable computational tool for bulk RNA-seq data decoding.

The rapid growth of biological data has intensified the need for transparent, reproducible, and well-documented computational workflows. The ability to clearly connect the steps of a workflow in the code with their description in a paper would improve workflow comprehension, support reproducibility, and facilitate reuse. This task requires the linking of bioinformatics tools in workflow code with their mentions in a published workflow description.