Abstract Motivation Single-cell RNA-seq captures tumor heterogeneity, but it is rarely available together with patient survival outcomes. Bulk RNA-seq cohorts provide survival data but lack direct measurements of cellular expression. Results DeSCENT generates reference-informed cellular profiles from bulk RNA-seq and combines both representations for survival prediction. We evaluated DeSCENT with three survival models on eight TCGA cancer cohorts. Compared with Bulk, DeSCENT improved average C-index and generally improved time-dependent AUC, while integrated Brier score results were mixed. External validation showed higher average C-index than all three baselines, although performance was lower than in internal evaluation. These findings support the use of cellular representations in the evaluated transcriptome-only setting. Availability and implementation Code: https://github.com/YonghaoZhao722/DeSCENT. Archival snapshot: doi:10.5281/zenodo.22073099. Supplementary information Supplementary data are available at Bioinformatics online.
Abstract Motivation Gene regulatory network (GRN) inference from single-cell transcriptomes is complicated by expression patterns associated with both cell state and shared ancestry. Cellular phylogenies provide developmental context, but incorporating relationships among individual cells into gene-level regulatory inference remains challenging. Results We present scPhyloGRN, a supervised framework that integrates single-cell gene expression with cellular phylogeny to prioritize directed transcription factor (TF)–target candidates within the observed gene set. The framework maps cell ancestry to gene associations that guide graph construction, attention in a dual-stream encoder, and lineage-consistency regularization. We evaluated scPhyloGRN against ten methods across 17 dataset–reference settings spanning four mouse embryos and one Caenorhabditis elegans dataset. scPhyloGRN showed an overall advantage in regulatory edge prioritization, supported by additional evaluations using shared candidate spaces. Removing lineage information reduced AUPRC in the ablation analyses, while branch-level and regulator-level analyses showed improved ranking and recovery of reference-supported regulatory candidates. Availability and implementation Source code and analysis workflows are available at https://github.com/Maria0852/scPhyloGRN, with a versioned archive at https://doi.org/10.5281/zenodo.22793831.
Abstract Motivation Explainable and accurate Survival Prediction (SP) remain an open challenge in Precision Oncology. Early SP methods achieved interpretability through simple models, often at the expenses of prediction performance. More recent Machine Learning (ML) methods can learn complex patterns, but they often resemble inscrutable black boxes. Here we propose SurvXAI, a new framework that combines reliable SP with the latest eXplainable AI (XAI) approach. Results We first train a Random Survival Forest model to predict the survival function of each patient. We then interpret these prediction by extracting patient-specific feature contributions with SurvSHAP, and we feed them into locally deployed Large Language Models (LLMs), to translate the numeric individual-level interpretation of the SP into natural language narrative explanations. Since conventional numerical breakdowns of the features contributions to the final predictions are often perceived as un-intuitive to non-ML experts, SurvXAI overcomes this limitation by automatically translating them into narrative explanations that can be readily used by clinicians, to facilitate the interpretation of model predictions. We validated our SP pipeline on six oncological datasets, including an in-house Multiple Myeloma cohort, comparing it with nine state of the art SP predictors. We then quantified the quality and usefulness of SurvXAI narrative explanations with a blind evaluation carried by seven human experts. This evaluation showed that larger LLMs produced more coherent explanations with accurate feature ranking and clinically plausible reasoning. SurvXAI demonstrates the potential of integrating explainable SP with LLMs to enable trustworthy, personalized interpretation in oncology and other clinical fields. Availability and Implementation SurvXAI is available at https://github.com/AI-for-Genome-Inter pretation-lab/SurvXAI_pipeline. The module for narrative generation is available at https://github.com/AI-for-Genome-Interpretation-lab/SurvX AI_stories.
Abstract Motivation Accurate Enzyme Commission (EC) number prediction is fundamental to protein function annotation and metabolic pathway reconstruction. Recent methods based on protein language model embeddings or 3D structure have improved EC prediction, but important limitations remain. Whole-sequence embeddings can dilute local functional signals, such as catalytic residues and functional domains, whereas structure-based methods depend on accurate protein structures and are therefore difficult to apply broadly. Results We present MIL-BC, a multiple instance learning framework for EC prediction that models each protein as a bag of domains rather than a single sequence-level object. This design is motivated by the observation that enzymatic function is often determined by constituent functional domains instead of the full-length sequence alone. Domain instances are generated automatically using our DCTdomain algorithm, avoiding dependence on curated domain annotations. MIL-BC aggregates domain-level evidence into protein-level EC predictions and assigns interpretable weights to domains that contribute most strongly to each prediction. We further combine the domain embedding with global embedding in MIL-BC+. MIL-BC+ achieved F1 improvements of 0.03 - 0.17 over the sequence-based baseline on all four test sets. The largest gains were observed for proteins with multiple enzymatic functions, where domain-level modeling better separates mixed functional signals. These results show that domain-centric multiple instance learning is an accurate and interpretable strategy for enzyme classification. By improving EC prediction without requiring structural data or curated domain labels, MIL-BC provides a scalable framework for large-scale protein function annotation. Availability Code is available as a GitHub repository at https://github.com/omics-ai/enzyme-mil.git.
Abstract Motivation Fine-grained and rare spatial domains are difficult to identify in spatial transcriptomics because they contain few spots, show subtle transcriptional differences, and are strongly influenced by neighboring regions. Results We present TRACER, a twin-contrastive learning framework that integrates a graph convolutional autoencoder with neighborhood-aware instance-level and cluster-level contrastive learning. By jointly preserving local spatial continuity and enhancing domain separability, TRACER achieved competitive and generally robust performance compared with representative spatial domain identification methods across multiple spatially resolved transcriptomic datasets, including human dorsolateral prefrontal cortex generated by 10× Visium, mouse brain datasets generated by STARmap, osmFISH, and BARISTAseq, and human breast cancer. TRACER recovered major cortical and anatomical structures, improved the identification of rare or transitional domains in the mouse somatosensory cortex, and revealed candidate functionally distinct small domains within a broadly annotated breast cancer region. Availability and implementation The source code of TRACER is available at https://github.com/luxin-heart/TRACER.
