Biological function is mediated by the hierarchical organization of cell types and states within tissue ecosystems. Identifying interpretable composite marker sets that both define and distinguish hierarchical cell identities is essential for decoding biological complexity yet remains a major challenge. Here, we present RECOMBINE, an algorithm that identifies recurrent composite marker sets to define hierarchical cell identities. Validation using both simulated and biological data sets demonstrates that RECOMBINE is robust to hyperparameter variation and data sparsity, and achieves higher accuracy in identifying discriminant markers compared with existing approaches. As a partition-free framework, RECOMBINE is particularly powerful for data sets characterized by continuous cell-state transitions, in which defining discrete boundaries is inappropriate. This capability is demonstrated by its application to zebrafish development, revealing gradual transcriptional transitions across embryonic stages, and to the mouse cerebellum, in which it uncovers transcriptional variation shaped by spatial gradients. When applied to single-cell data and validated with spatial transcriptomic data from the mouse visual cortex, RECOMBINE identifies key cell-type markers and generates a robust gene panel for targeted spatial profiling. It also uncovers markers of CD8+ T cell states, including GZMK+HAVCR2- effector memory cells associated with anti-PD-1 therapy response. Finally, using data from the Tabula Sapiens project, RECOMBINE identifies composite marker sets across a broad range of human tissues. Together, these results highlight RECOMBINE as a robust, data-driven framework for optimized marker selection, enabling the discovery and validation of hierarchical cell identities across diverse tissue contexts.
Advances in high-throughput sequencing have lowered the cost and complexity of genome sequencing, making it possible for the first time to assemble large pangenomic data sets for many species. These data sets, comprising thousands of individuals, already span from hundreds of gigabytes to petabytes, far exceeding the memory capacity of most machines, and are expected to continue growing in scale over time. Already, many traditional bioinformatics tools fail on inputs at this scale because they cannot construct their necessary data structures within memory limits. There is a growing need for methods that can construct these structures directly from compressed representations. Prefix-free parsing (PFP) addresses this challenge. PFP serves as a preprocessing step that compresses sufficiently repetitive text, yet still permits building important data structures for the original data set from its compressed output. This survey offers an overview of PFP, covering its core principles, the primary data structures it enables, current applications, and future research directions.
Gene transcription is activated through the interaction between cis-regulatory elements (CREs) and transcription factors (TFs). CREs serve as templates to provide binding sites, whereas TFs provide functions to directly initiate transcription. Current research mainly focuses on deciphering cis-regulatory code, while neglecting TF regulatory code. However, CREs alone are not sufficient to determine the binding of TFs, which makes the interpretation of cis-regulatory code ambiguous. In this study, we systematically analyze 13 TF binding profiles associated with transcription initiation and encode them as TF sequence to explore the TF regulatory code. Furthermore, we propose a deep learning model named DeepTF to predict gene expression from TF sequence. Results show that TF binding exhibits conserved positional preferences and combinatorial patterns in promoters and DeepTF is able to predict gene expression with high accuracy (AUROC = 0.97). Meanwhile, cross-cell-line validation (AUROC > 0.90) further confirms the model's transferability. Model interpretation reveals DeepTF successfully captures the TF regulatory grammar associated with gene expression. Compared with the cis-regulatory code, our proposed TF regulatory code is better suited for investigating the relationship between TF binding positions or combinatorial patterns and gene expression. Collectively, DeepTF simplifies gene expression prediction and provides clear biological insights of transcriptional regulation.
Cytosine methylation, a crucial epigenetic modification, plays a vital role in genomic regulation. Leveraging the advancements in long-read sequencing, we investigate the methylation patterns of polymorphic transposable element (TE) insertions of human lymphoblastoid cell lines (LCLs). We validate the high concordance between long-read methylation calls and the conventional whole-genome bisulfite sequencing (WGBS) method. We then aim to establish general rules of TE methylation with our data by addressing three key questions: (1) what is the methylation profile of each insertion; (2) do newly inserted TEs adopt the methylation pattern of their genomic context; and (3) do new TE insertions affect the methylation of their flanking regions. Although most non-TE insertions exhibit DNA methylation patterns consistent with their genomic context, TE insertions are generally highly methylated, exhibiting distinct, class-specific patterns with some variation within TE bodies. A small percentage of Alu insertions are hypomethylated, particularly those inserted within hypomethylated CpG islands. We also reveal that majority of TEs exhibited minimal impact on nearby regions, although numerous exceptions exist in which the methylation status of TEs spread into nearby regions. In conclusion, although TE insertions primarily exhibit methylation patterns restricted within their boundaries, some TEs are able to affect the methylation level of their genomic neighborhoods. At last, our findings are limited to human LCLs, and more comprehensive analysis would be needed to test the rules we found here on a broader spectrum of cell types and developmental stages.
Precise transcription factor (TF) binding to DNA governs gene regulation, yet nucleotide sequence alone often fails to fully capture binding specificity. Although static DNA shape is a recognized determinant of indirect readout, the role of intrinsic conformational flexibility remains underexplored across TF families. Here, we demonstrate that integrating sequence-derived DNA flexibility descriptors into predictive models improves both prediction and mechanistic interpretability of TF-DNA affinity. Across large-scale in vitro data sets encompassing HT-SELEX and protein-binding microarrays for mammalian and Drosophila TFs, flexibility-augmented models consistently outperform sequence-only baselines and complement DNA shape models. Cross-platform analyses further indicate that flexibility features capture structural information that is robust to platform-specific biases. Using a position-resolved interpretation framework, we uncover family-specific "flexibility footprints," including recurrent hotspots in core motifs and flanks that align with DNA structural deformations from TF-DNA complex structures. Extending to ENCODE ChIP-seq and DNase-seq data, flexibility augmentation improves classification of functional TF binding sites across diverse TFs and cellular contexts. Collectively, these results underscore the insufficiency of sequence-only models and highlight the utility of the flexibility descriptors as an interpretable component of the TF recognition code.
