Pi-Jing Wei

The precise identification of promoters is crucial for understanding gene regulation. Deep learning methods have achieved considerable success in promoter prediction, yet most operate at the sequence level with coarse-grained labels. This means they label an entire DNA segment as either a “promoter” or “non-promoter,” which results in a lack of the nucleotide-level resolution in prediction. In this study, we propose EvoSNR-Prom, a model designed for promoter prediction at single-nucleotide resolution. EvoSNR-Prom is built on the Evo foundation model and formulates promoter identification as a token-level sequence labeling problem, analogous to named entity recognition in natural language processing. To address the limited contextual information available in single-nucleotide tokenization, we introduce a lexicon-enhanced embedding strategy that incorporates biologically meaningful DNA lexicons, enriching contextual representations and improving the model’s ability to capture complex sequence motifs. Furthermore, to enhance predictive performance on small size datasets, we integrate a label-aware transfer learning framework to leverage knowledge from well-annotated source species to a target organism. The results across various prokaryotic datasets show that EvoSNR-Prom achieves excellent performance. This work provides a valuable computational framework for the high-precision analysis of gene regulatory elements, contributing to the advancement of promoter prediction at single-nucleotide resolution.

G-quadruplexes (G4s) are non-canonical nucleic acid secondary structures that help maintain genomic stability and regulate gene transcription. Although the genome contains a vast number of putative G4-forming sequences (PQSs, sequences with intrinsic in vitro G4-forming potential), only a small fraction fold stably into G4 structures within the complex chromatin environment of living cells. Existing deep learning approaches improve predictive accuracy by incorporating cell line–specific epigenetic data; however, their heavy reliance on costly, large-scale sequencing assays (e.g., ChIP-seq) limits broader application to clinical samples and newly profiled cell lines. To address this challenge, we propose iDualG4, an interpretable dual-channel deep learning framework that uses DNA sequence as the only input at inference time. By leveraging a pretrained Enformer module, iDualG4 infers epigenomic proxy features directly from DNA sequence and integrates them with local sequence features, thereby replacing the need for newly measured cell-specific epigenomic assays during prediction. Evaluations across multiple cell lines, including K562, demonstrate that iDualG4 significantly outperforms existing methods, particularly in handling imbalanced data (achieving an AUPR of 0.981 on K562). Interpretability analysis based on DeepSHAP indicates that iDualG4 provides an in vivo G4 prediction tool combining high precision and interpretability without the need for additional experimental sequencing data, and offers a novel computational framework for elucidating how sequence and the epigenetic environment jointly determine genomic G4 formation.

The radiosensitivity of cancer patients determines the efficacy of radiotherapy, and patients with low radiosensitivity cannot benefit from radiotherapy. Therefore, accurately predicting radiosensitivity before treatment is essential for personalized and precise radiotherapy. However, most existing studies rely solely on genomic and clinical features, neglecting the tumor microenvironmental information embedded in histopathological images, which limits prediction accuracy. To address this issue, we propose Resfusion, a deep multimodal fusion framework that integrates patient-level gene expression profiles, clinical records, and histopathological images for tumor radiosensitivity prediction. Specifically, the pre-trained large-scale pathology model is used as an image encoder to extract global representations from whole-slide pathological image. Radiosensitivity-related genes are selected using an autoencoder combined with univariate Cox regression, while clinically relevant variables are manually curated. The three modalities are first concatenated and then refined through a self-attention-based module, which captures inter-feature dependencies within the fused representation and highlights complementary information across modalities. The model was evaluated using five-fold cross-validation on two common tumor datasets suitable for radiotherapy: the Breast Invasive Carcinoma (BRCA) dataset (282 patients in total, with each fold partitioned into 226 training samples and 56 validation samples) and the Head and Neck Squamous Cell Carcinoma (HNSC) dataset (200 patients in total, with each fold partitioned into 161 training samples and 39 validation samples). The average AUC values obtained from the five-fold cross-validation reached 76.83% and 79.49%, respectively. Experimental results demonstrate that the Resfusion model significantly outperforms unimodal methods and existing multimodal fusion methods, verifying its effectiveness in predicting the radiosensitivity of tumor patients.

Graph transformer networks have received more attention in hyperspectral image (HSI) classification. However, they overlooked the influence of graph connectivity strength in positional encoding and distribution. In order to address the above deficiencies, we proposed the novel graph transformer with structural embedding and training (GTSET) for HSI classification. Specifically, the structural embedding module firstly aimed at extracting effectively local and non-local feature information via patch-based distance encoding and centrality correlation coefficients based on graph connectivity strength, alleviating spectral variability. Secondly, the structural training module aimed at addressing imbalanced structural position distribution of labeled samples by leveraging the topological graph connectivity to determine their structural position distribution and reweighting the influence of labeled samples on the graph transformer training stage, exploring the guiding role of labeled samples in low spatial resolution of HSI. Next, we further refine training weights based on the spectral feature smoothness of labeled samples. Finally, comprehensive experiments on three real-world HSI datasets demonstrate that the GTSET achieves superior performance in HSI classification with limited labeled samples, compared to other popular classification methods. Implementation of GTSET, along with examples, can be found on the GitHub repository: https://github.com/xuchengchao0/GTSET.

Promoters are core elements in regulating gene expression. The design and optimization of functional promoters is crucial for enhancing metabolic pathway construction and advancing gene therapy. Deep learning-based methods have shown great potential in promoter design. However, existing studies mainly focus on designing strong promoters, neglecting the practical need for promoters with varying regulatory intensities. Here, we propose a novel promoter design method, PromoDGDE, to design promoters with desirable expression levels and apply it to the promoter design of Escherichia coli and Saccharomyces cerevisiae. It uses Diffusion-GAN to learn the feature distribution of natural sequences and generate new promoters. Then, reinforcement learning and evolutionary algorithms are combined to dynamically optimize the synthetic sequences. In silico analyze results demonstrate that PromoDGDE outperforms existing methods, generating promoters that not only possess biological significance but also achieve the intended function. In vivo experiment results demonstrate that the synthetic promoters exhibit expression activity, with over 60% of the sequences showing the expected regulatory effects. These results confirm the practical effectiveness of PromoDGDE and demonstrate its ability to provide an efficient and flexible solution for complex design needs in synthetic biology.

Reconstructing high-resolution gene regulatory networks (GRNs) based on single-cell RNA sequencing data provides an opportunity to gain insight into disease pathogenesis. At present, there are a large number of GRN reconstruction methods based on graph neural networks, and they can obtain excellent performance in GRN inference by extracting network structure features. However, most of these methods fail to fully exploit the directional characteristics or even ignore them when extracting network structural features. To this end, a novel framework called GAEDGRN is proposed based on gravity-inspired graph autoencoder (GIGAE) to infer potential causal relationships between genes. Among them, GIGAE can help us capture the complex directed network topology in GRN. Additionally, due to the uneven distribution of the latent vectors generated by the graph autoencoder, a random walk-based method is used to regularize the latent vectors learnt by the encoder. Furthermore, considering that some genes in GRN usually have a significant impact on biological functions, GAEDGRN designs a gene importance score calculation method and pays attention to genes with high importance in the process of GRN reconstruction. Experimental results on seven cell types of three GRN types show that GAEDGRN achieves high accuracy and strong robustness. Moreover, a case study on human embryonic stem cells demonstrates that GAEDGRN can help identify important genes.