Biology subjects

Zhou, S. K.

Publications and source records attributed to Zhou, S. K..

7 recordsLinked to original sources

bioRxiv2026

LigRNA: High-accuracy prediction of RNA-ligand binding sites via dynamic chemical complementarity and language model encoding

RNA-binding small molecules represent a validated therapeutic modality, yet pinpointing their exact RNA binding sites remains critical for mechanistic understanding and drug optimization. Experimental methods, while precise, are costly and low-throughput, necessitating robust computational alternatives. Approaches based on static feature concatenation may overlook input-dependent relationships between RNA residues and ligand functional groups. Here we present LigRNA, a graph neural network framework that predicts RNA-ligand binding sites through joint cross-modal encoding of both molecules. LigRNA integrates structure-guided language model embeddings with physicochemical features-solvent-accessible surface area and secondary structure types, to construct comprehensive RNA representations. Ligands are encoded as attributed molecular graphs via graph attention networks, preserving atom-level chemical topology rather than compressing molecules into fixed fingerprints. A cross-attention mechanism then enables dynamic, ligand-dependent modulation of RNA nucleotide attention, capturing context-specific interaction patterns. Benchmarked on TE9 and TE18 datasets, LigRNA achieves F1 scores of 0.73 and 0.58, improving on GATRsite by 7.4% and 18.4%, respectively. Ablation studies identify the RNA residual encoder as the most consistent contributor across the two benchmarks. The source code is freely available at https://github.com/WillowWang0216/LigRNA.

bioinformatics↗

bioRxiv2026

ReTIF: Granularity-Aware Multitask Interaction Routing for RNA-Compound Interaction Prediction and Binding-Site Localization

RNA-targeted drug discovery requires both RNA-compound interaction prediction and nucleotide-level binding-site (BS) localization. These two tasks rely on different levels of interaction information: DTI prediction summarizes overall RNA-compound compatibility, whereas BS localization requires preserving nucleotide-level compound-associated signals. However, existing multitask models often use shared cross-modal representations for both tasks before task-specific prediction layers, which may limit their ability to preserve task-dependent interaction patterns. We therefore propose ReTIF (Relation-enhanced Task-specific Interaction Framework), which constructs separate cross-modal interaction representations for DTI prediction and BS localization before aggregation. ReTIF integrates multi-source RNA-compound representations from frozen RNA-FM, StructRFM, Mole-BERT, and MolFormer encoders, and builds separate interaction representations for DTI scoring and BS localization. The DTI branch captures global compatibility through aggregation, whereas the BS branch preserves nucleotide-compound resolution and enhances local evidence through relation-guided propagation with RNA structural and compound topological priors. An asymmetric DTI-derived compatibility signal provides global context to BS prediction while maintaining local evidence. Across five-fold evaluations under unseen pair, RNA, compound, and joint shifts with 13 baselines, ReTIF has the highest mean in 13 of 16 scenario-metric combinations, including BS AUPR in all four settings.

bioinformatics↗

bioRxiv2026

H2O: A Foundation Model Bridging Histopathology to Spatial Multi-Omics Profiling

Spatial omics technologies have revolutionized the molecular profiling of tissues but remain constrained by high costs and limited scalability. While hematoxylin and eosin (H&E) staining is ubiquitous, it lacks molecular specificity. Here, we present H2O (Histopathology to Omics), a generalist AI framework that bridges the modality gap between histopathology and spatial multi-omics, enabling the direct inference of spatial transcriptomics (ST) and proteomics (SP) landscapes from routine H&E images. H2O integrates Vision Transformers (ViT) with Large Language Models (LLM) via contrastive learning to align histological morphology with semantic molecular knowledge. This cross-modal approach allows the model to incorporate spatial expression profiles into histological pattern recognition, effectively decoding the molecular heterogeneity underlying tissue morphology. Trained on a pan-tissue dataset of 1.3 million paired H&E-spatial patches across 25 organs and cancer types, H2O predicts spatial omics expression from histology with high concordance to sequenced measurements and consistently outperforms state-of-the-art models across three cancer benchmarks. Notably, H2O recovers the MIF-CD74/CD44 signaling axis directly from H&E images, highlighting its capacity to infer biologically meaningful cell-cell communication without molecular profiling. Applying on three additional public cohorts covering fetal and paediatric thymus tissues, human metastatic lymph node, and breast cancer, encompassing human development, 3D spatial frameworks, and integrative multi-omics, H2O yields biologically concordant insights, demonstrating superior accuracy, robustness, and generalizability across real-world applications in diverse scenarios. H2O converts routine histopathology into a portal for spatially resolved multi-omics profiling by computationally generating transcriptomic and proteomic landscapes, thereby enhancing tissue phenotyping and enabling scalable, integrative tissue-atlas construction.

bioinformatics↗

bioRxiv2025

Toward Accurate RNA Non-Canonical Structure Prediction: The NC-Bench Benchmark and the NCfold Framework

AO_SCPLOWBSTRACTC_SCPLOWAccurate prediction of non-canonical (NC) base pairs is a pivotal step toward un-raveling the full functional landscape of RNA. This goal requires both a standardized dataset for evaluation and powerful models that overcome data limitations. Here, we make a dual contribution to this endeavor: the NC-Bench benchmark and the NCfold framework. NC-Bench offers a trusted ground for comparison with 925 sequences and 6,708 annotated NC pairs, defining tasks for edge and orientation prediction. NCfold provides a powerful analytical engine, a closed-loop dual-branch network that synergizes sequence features with structure-aware embeddings from RNA foundation models via our proposed REFweighted self-attention. This iterative design enables robust learning from scarce NC data. On the NC-Bench benchmark, NCfold establishes new state-of-the-art results, significantly outperforming existing methods. Comprehensive analyses validate its design and affirm the critical role of NC-Bench in driving future progress. NC-Bench and NCfold together form a cornerstone for the next generation of RNA structure prediction. The datasets and codes are publicly available at https://github.com/heqin-zhu/NCBench

bioinformatics↗

bioRxiv2025

A fully-open structure-guided RNA foundation model for robust structural and functional inference

RNA language models have achieved strong performances across diverse downstream tasks by leveraging large-scale sequence data. However, RNA function is fundamentally shaped by its hierarchical structure, making the integration of structural information into pre-training essential. Existing methods often depend on noisy structural annotations or introduce task-specific biases, limiting model generalizability. Here, we propose structRFM, a structure-guided RNA foundation model that is pre-trained on millions of RNA sequences and secondary structures data by integrating base pairing interactions into masked language modeling through a novel pair matching operation. We further introduce MUSES (multi-source ensemble of secondary structures) to mitigate model bias, and a dynamic masking ratio to balance the structure-guided mask and nucleotide-level mask. structRFM learns joint knowledge of sequential and structural data, producing versatile representations, including classification-level, sequence-level, and pairwise matrix features, that support a broad spectrum of downstream adaptations. structRFM ranks among the top models in zero-shot homology classification across seventeen biological language models, and sets new benchmarks for secondary structure prediction. structRFM further derives Zfold, which enables robust and reliable tertiary structure prediction, with consistent improvements in estimating 3D structures and their accordingly extracted 2D structures, achieving a pronounced about 20% performance gain compared with baselines and comparable performances with AlphaFold3 on CASP15-natural, CASP16, and RNA-Puzzles datasets. In functional tasks such as internal ribosome entry site identification, structRFM achieves a whopping 48% performance gain in F1 score. Furthermore, state-of-the-art performances in extensive experiments across novel RNA families and long non-coding RNAs indicate the robustness and generalizability of structRFM. These results demonstrate the effectiveness of structure-guided pre-training and highlight a promising direction for developing multi-modal RNA language models in computational biology. To support the broader scientific community, we have made the 21-million sequence-structure dataset and the pre-trained structRFM model fully open-source, facilitating the development of multimodal foundation models in biology.

bioinformatics↗

bioRxiv2024

Deep generalizable prediction of RNA secondary structure via base pair motif energy

Deep learning methods have demonstrated great performance for RNA secondary structure prediction. However, generalizability is a common unsolved issue on unseen out-of-distribution RNA families, which hinders further improvement of the accuracy and robustness of deep learning methods. Here we construct a base pair motif library that enumerates the complete space of the locally adjacent three-neighbor base pair and records the thermodynamic energy of corresponding base pair motifs through de novo modeling of tertiary structures, and we further develop a deep learning approach for RNA secondary structure prediction, named BPfold, which learns relationship between RNA sequence and the energy map of base pair motif. Experiments on sequence-wise and family-wise datasets have demonstrated the great superiority of BPfold compared to other state-of-the-art approaches in accuracy and generalizability. We hope this work contributes to integrating physical priors and deep learning methods for the further discovery of RNA structures and functionalities.

bioinformatics↗

bioRxiv2023

A systematic search for RNA structural switches across the human transcriptome

RNA structural switches are key regulators of gene expression in bacteria, yet their characterization in Metazoa remains limited. Here we present SwitchSeeker, a comprehensive computational and experimental approach for systematic identification of functional RNA structural switches. We applied SwitchSeeker to the human transcriptome and identified 245 putative RNA switches. To validate our approach, we characterized a previously unknown RNA switch in the 3UTR of the RORC transcript. In vivo DMS-MaPseq, coupled with cryogenic electron microscopy, confirmed its existence as two alternative structural conformations. Furthermore, we used genome-scale CRISPR screens to identify trans factors that regulate gene expression through this RNA structural switch. We found that nonsense-mediated mRNA decay acts on this element in a conformation-specific manner. SwitchSeeker provides an unbiased, experimentally-driven method for discovering RNA structural switches that shape the eukaryotic gene expression landscape.

systems biology↗