Single-cell multi-omics technologies provide complementary molecular information for characterizing cellular heterogeneity. However, sparsity, noise, cross-omics heterogeneity, and imbalanced omics information make accurate cell clustering challenging. We propose scCIF, an end-to-end contribution-informed fusion framework for single-cell multi-omics clustering. scCIF employs contribution-aware fusion to adaptively construct a consensus semantic space, and further introduces hard-aware assignment calibration to align omics-specific assignments with the fused assignment distribution. Experiments on eight public single-cell multi-omics datasets show that scCIF consistently outperforms competing methods across ARI, NMI, and AMI. Ablation studies verify the effectiveness of each key component. UMAP visualization, batch effect analysis, trajectory inference, and marker gene analysis further demonstrate that scCIF preserves biologically meaningful cellular structures. scCIF improves single-cell multi-omics clustering by jointly modeling contribution imbalance and cross-omics semantic drift, providing a robust framework for learning discriminative and biologically interpretable cell representations.
While bulk RNA sequencing offers clinical practicality through cost-efficiency and scalability for large cohort profiling, it obscures cellular heterogeneity critical for characterizing tissue microenvironment. Conversely, single-cell RNA sequencing (scRNA-seq) resolves cellular diversity at unprecedented resolution but faces translational barriers including technical complexity and tissue processing artifacts. These complementary yet constrained technologies necessitate computational frameworks that leverage scRNA-seq data insights to enhance bulk data interpretation. Current deconvolution methods, however, remain challenged by technical noise amplification, biological implausibility, and limited rare-cell detection across diverse platforms. We developed ESMF (Elastic Signature-based Matrix Factorization), a deconvolution framework addressing these gaps through: (1) elastic expression signature-based regularization incorporating gene expression distribution parameters (mean/variance) from scRNA-seq data references, (2) negative binomial modeling of technical noise, and (3) dual-validation of marker selection integrating literature-curated knowledge with computational screening. Validated across diverse synthetic and biological datasets using a multi-tiered framework, ESMF demonstrated competitive performance against established benchmarks, exhibiting improved cross-platform stability, enhanced rare-cell detection in clinically relevant samples, and biologically plausible expression estimates. ESMF effectively bridges bulk and single-cell paradigms by resolving cellular proportions and expression profiles with biological fidelity, providing a robust analytical platform for translational tissue microenvironment investigation.
Single-cell RNA sequencing enables high-resolution characterization of cellular transcriptomes, but learning robust representations remains challenging because of sparsity, dropout, and technical noise. Methods that rely directly on raw gene expression are highly sensitive to these artifacts, even though they can capture dataset-specific transcriptional patterns in individual datasets. In contrast, large-scale pretrained single-cell models provide global biological priors through pretrained gene representations, but often fail to capture context-dependent signals. As a result, methods that effectively integrate pretrained biological priors with dataset-specific expression patterns remain relatively limited. We present scDECA, a dual-encoder cross-attention framework that integrates raw gene expression, pretrained gene priors, and local structure derived from protein–protein interaction networks and expression-based cell similarity graphs. The gene encoder incorporates raw expression, pretrained gene embeddings, and protein–protein interaction structure, while the cell encoder models context-specific cellular relationships using a similarity graph constructed from observed expression profiles. Cross-attention aligns the gene and cell domains to produce unified representations that capture both global biological organization and dataset-specific transcriptional heterogeneity. Across multiple disease datasets, scDECA improves clustering performance, recovers functional gene programs more consistently, and produces biologically interpretable representations. These results demonstrate the benefit of jointly modeling global biological priors and local structure for robust and biologically meaningful representation learning in single-cell transcriptomics. scDECA provides an effective framework for integrating pretrained knowledge with dataset-specific transcriptional structure, improving interpretability and downstream analysis performance.
Learning robust and discriminative representations from gastrointestinal endoscopy images remains challenging due to high visual variability, subtle pathological differences, and pronounced class imbalance. While variational autoencoders (VAEs) have shown promise for unsupervised and semi-supervised representation learning, directly constructing latent spaces from high-dimensional deep features often leads to redundant and clinically irrelevant representations. In this study, we introduce a deep feature learning approach that is called EvoLatent-FS, led by evolution, that directly optimizes feature representations before latent modeling. The suggested method incorporates an intermediary evolutionary optimization stage that refines deep feature representations prior to variational encoding, in contrast to current approaches that either perform feature selection after representation learning or apply latent learning directly to raw deep features. This architecture preserves diagnostically relevant patterns while suppressing superfluous visual input, allowing the learning of more structured and discriminative latent spaces. The three primary parts of the framework are latent representation learning using a variational autoencoder, population-based evolutionary feature optimization via binary feature masking, and deep feature extraction using a convolutional backbone. The suggested method creates a unique representation learning pipeline that is suited to the difficulties of gastrointestinal endoscopic imaging by separating feature optimization from latent modeling and putting it directly before variational encoding. The suggested framework produces more compact latent representations and better downstream classification performance than traditional pipelines that directly apply latent modeling to unoptimized deep features, according to extensive experiments carried out on labeled gastrointestinal endoscopy image datasets. These findings demonstrate the efficacy of evolutionary-guided pre-latent feature optimization as a novel and principled approach to medical image representation learning.
