Cell reprogramming, which guides the conversion between cell states, is a promising technology for tissue repair and regeneration, with the ultimate goal of accelerating recovery from diseases or injuries. To accomplish this, regulators must be identified and manipulated to control cell fate. We propose Fatecode, a computational method that predicts cell fate regulators based only on single-cell RNA sequencing (scRNA-seq) data. Fatecode learns a latent representation of the scRNA-seq data using a deep learning-based classification-supervised autoencoder and then performs in silico perturbation experiments on the latent representation to predict genes that, when perturbed, would alter the original cell type distribution to increase or decrease the population size of a cell type of interest. We assessed Fatecode\'s performance using simulations from a mechanistic gene-regulatory network model and scRNA-seq data mapping blood and brain development of different organisms. Our results suggest that Fatecode can detect known cell fate regulators from single-cell transcriptomics datasets.

In single-cell RNA sequencing (scRNA-seq) studies, cell types and their marker genes are often identified by clustering and differentially expressed gene (DEG) analysis. A common practice is to select genes using surrogate criteria such as variance and deviance, then cluster them using selected genes and detect markers by DEG analysis assuming known cell types. The surrogate criteria can miss important genes or select unimportant genes, while DEG analysis has the selection-bias problem. We present Festem, a statistical method for the direct selection of cell-type markers for downstream clustering. Festem distinguishes marker genes with heterogeneous distribution across cells that are cluster informative. Simulation and scRNA-seq applications demonstrate that Festem can sensitively select markers with high precision and enables the identification of cell types often missed by other methods. In a large intrahepatic cholangiocarcinoma dataset, we identify diverse CD8+ T cell types and potential prognostic marker genes.

Deep-learning tools that extract prognostic factors derived from multi-omics data have recently contributed to individualized predictions of survival outcomes. However, the limited size of integrated omics-imaging-clinical datasets poses challenges. Here, we propose two biologically interpretable and robust deep-learning architectures for survival prediction of non-small cell lung cancer (NSCLC) patients, learning simultaneously from computed tomography (CT) scan images, gene expression data, and clinical information. The proposed models integrate patient-specific clinical, transcriptomic, and imaging data and incorporate Kyoto Encyclopedia of Genes and Genomes (KEGG) and Reactome pathway information, adding biological knowledge within the learning process to extract prognostic gene biomarkers and molecular pathways. While both models accurately stratify patients in high- and low-risk groups when trained on a dataset of only 130 patients, introducing a cross-attention mechanism in a sparse autoencoder significantly improves the performance, highlighting tumor regions and NSCLC-related genes as potential biomarkers and thus offering a significant methodological advancement when learning from small imaging-omics-clinical samples.

Gene co-expression analysis of single-cell transcriptomes, aiming to define functional relationships between genes, is challenging due to excessive dropout values. Here, we developed a single-cell graphical Gaussian model (SingleCellGGM) algorithm to conduct single-cell gene co-expression network analysis. When applied to mouse single-cell datasets, SingleCellGGM constructed networks from which gene co-expression modules with highly significant functional enrichment were identified. We considered the modules as gene expression programs (GEPs). These GEPs enable direct cell-type annotation of individual cells without cell clustering, and they are enriched with genes required for the functions of the corresponding cells, sometimes at levels greater than 10-fold. The GEPs are conserved across datasets and enable universal cell-type label transfer across different studies. We also proposed a dimension-reduction method through averaging by GEPs for single-cell analysis, enhancing the interpretability of results. Thus, SingleCellGGM offers a unique GEP-based perspective to analyze single-cell transcriptomes and reveals biological insights shared by different single-cell datasets.

The cellular components of tumors and their microenvironment play pivotal roles in tumor progression, patient survival, and the response to cancer treatments. Unveiling a comprehensive cellular profile within bulk tumors via single-cell RNA sequencing (scRNA-seq) data is crucial, as it unveils intrinsic tumor cellular traits that elude identification through conventional cancer subtyping methods. Our contribution, scBeacon, is a tool that derives cell-type signatures by integrating and clustering multiple scRNA-seq datasets to extract signatures for deconvolving unrelated tumor datasets on bulk samples. Through the employment of scBeacon on the The Cancer Genome Atlas (TCGA) cohort, we find cellular and molecular attributes within specific tumor categories, many with patient outcome relevance. We developed a tumor cell-type map to visually depict the relationships among TCGA samples based on the cell-type inferences.

Cancer of unknown primary (CUP) represents metastatic cancer where the primary site remains unidentified despite standard diagnostic procedures. To determine the tumor origin in such cases, we developed BPformer, a deep learning method integrating the transformer model with prior knowledge of biological pathways. Trained on transcriptomes from 10,410 primary tumors across 32 cancer types, BPformer achieved remarkable accuracy rates of 94%, 92%, and 89% in primary tumors and primary and metastatic sites of metastatic tumors, respectively, surpassing existing methods. Additionally, BPformer was validated in a retrospective study, demonstrating consistency with tumor sites diagnosed through immunohistochemistry and histopathology. Furthermore, BPformer was able to rank pathways based on their contribution to tumor origin identification, which helped to classify oncogenic signaling pathways into those that are highly conservative among different cancers versus those that are highly variable depending on their origins.

Plasma cell-free DNA (cfDNA) fragmentation patterns are emerging directions in cancer liquid biopsy with high translational significance. Conventionally, the cfDNA sequencing reads are aligned to a reference genome to extract their fragmentomic features. In this study, through cfDNA fragmentomics profiling using different reference genomes on the same datasets in parallel, we report systematic biases in such conventional reference-based approaches. The biases in cfDNA fragmentomic features vary among races in a sample-dependent manner and therefore might adversely affect the performances of cancer diagnosis assays across multiple clinical centers. In addition, to circumvent the analytical biases, we develop Freefly, a reference-free approach for cfDNA fragmentomics profiling. Freefly runs ∼60-fold faster than the conventional reference-based approach while generating highly consistent results. Moreover, cfDNA fragmentomic features reported by Freefly can be directly used for cancer diagnosis. Hence, Freefly possesses translational merit toward the rapid and unbiased measurement of cfDNA fragmentomics.

Single-cell RNA sequencing (scRNA-seq) has transformed our understanding of cellular responses to perturbations such as therapeutic interventions and vaccines. Gene relevance to such perturbations is often assessed through differential expression analysis (DEA), which offers a one-dimensional view of the transcriptomic landscape. This method potentially overlooks genes with modest expression changes but profound downstream effects and is susceptible to false positives. We present GENIX (gene expression network importance examination), a computational framework that transcends DEA by constructing gene association networks and employing a network-based comparative model to identify topological signature genes. We benchmark GENIX using both synthetic and experimental datasets, including analysis of influenza vaccine-induced immune responses in peripheral blood mononuclear cells (PBMCs) from recovered COVID-19 patients. GENIX successfully emulates key characteristics of biological networks and reveals signature genes that are missed by classical DEA, thereby broadening the scope of target gene discovery in precision medicine.

We present an innovative strategy for integrating whole-genome-wide multi-omics data, which facilitates adaptive amalgamation by leveraging hidden layer features derived from high-dimensional omics data through a multi-task encoder. Empirical evaluations on eight benchmark cancer datasets substantiated that our proposed framework outstripped the comparative algorithms in cancer subtyping, delivering superior subtyping outcomes. Building upon these subtyping results, we establish a robust pipeline for identifying whole-genome-wide biomarkers, unearthing 195 significant biomarkers. Furthermore, we conduct an exhaustive analysis to assess the importance of each omic and non-coding region features at the whole-genome-wide level during cancer subtyping. Our investigation shows that both omics and non-coding region features substantially impact cancer development and survival prognosis. This study emphasizes the potential and practical implications of integrating genome-wide data in cancer research, demonstrating the potency of comprehensive genomic characterization. Additionally, our findings offer insightful perspectives for multi-omics analysis employing deep learning methodologies.

Predicting cellular responses to perturbations requires interpretable insights into molecular regulatory dynamics to perform reliable cell fate control, despite the confounding non-linearity of the underlying interactions. There is a growing interest in developing machine learning-based perturbation response prediction models to handle the non-linearity of perturbation data, but their interpretation in terms of molecular regulatory dynamics remains a challenge. Alternatively, for meaningful biological interpretation, logical network models such as Boolean networks are widely used in systems biology to represent intracellular molecular regulation. However, determining the appropriate regulatory logic of large-scale networks remains an obstacle due to the high-dimensional and discontinuous search space. To tackle these challenges, we present a scalable derivative-free optimizer trained by meta-reinforcement learning for Boolean network models. The logical network model optimized by the trained optimizer successfully predicts anti-cancer drug responses of cancer cell lines, while simultaneously providing insight into their underlying molecular regulatory mechanisms.