scDAC: deep adaptive clustering of single-cell transcriptomic data with coupled autoencoder and Dirichlet process mixture model.

MOTIVATION: Clustering analysis for single-cell RNA sequencing (scRNA-seq) data is an important step in revealing cellular heterogeneity. Many clustering methods have been proposed to discover heterogenous cell types from scRNA-seq data. However, adaptive clustering with accurate cluster number reflecting intrinsic biology nature from large-scale scRNA-seq data remains quite challenging. RESULTS: Here, we propose a single-cell Deep Adaptive Clustering (scDAC) model by coupling the Autoencoder (AE) and the Dirichlet Process Mixture Model (DPMM). By jointly optimizing the model parameters of AE and DPMM, scDAC achieves adaptive clustering with accurate cluster numbers on scRNA-seq data. We verify the performance of scDAC on five subsampled datasets with different numbers of cell types and compare it with 15 widely used clustering methods across nine scRNA-seq datasets. Our results demonstrate that scDAC can adaptively find accurate numbers of cell types or subtypes and outperforms other methods. Moreover, the performance of scDAC is robust to hyperparameter changes. AVAILABILITY AND IMPLEMENTATION: The scDAC is implemented in Python. The source code is available at https://github.com/labomics/scDAC.

Mosaic integration and knowledge transfer of single-cell multimodal data with MIDAS.

Integrating single-cell datasets produced by multiple omics technologies is essential for defining cellular heterogeneity. Mosaic integration, in which different datasets share only some of the measured modalities, poses major challenges, particularly regarding modality alignment and batch effect removal. Here, we present a deep probabilistic framework for the mosaic integration and knowledge transfer (MIDAS) of single-cell multimodal data. MIDAS simultaneously achieves dimensionality reduction, imputation and batch correction of mosaic data by using self-supervised modality alignment and information-theoretic latent disentanglement. We demonstrate its superiority to 19 other methods and reliability by evaluating its performance in trimodal and mosaic integration tasks. We also constructed a single-cell trimodal atlas of human peripheral blood mononuclear cells and tailored transfer learning and reciprocal reference mapping schemes to enable flexible and accurate knowledge transfer from the atlas to new data. Applications in mosaic integration, pseudotime analysis and cross-tissue knowledge transfer on bone marrow mosaic datasets demonstrate the versatility and superiority of MIDAS. MIDAS is available at https://github.com/labomics/midas .

Towards Strain-Level Complexity: Sequencing Depth Required for Comprehensive Single-Nucleotide Polymorphism Analysis of the Human Gut Microbiome.

Intestinal bacteria strains play crucial roles in maintaining host health. Researchers have increasingly recognized the importance of strain-level analysis in metagenomic studies. Many analysis tools and several cutting-edge sequencing techniques like single cell sequencing have been proposed to decipher strains in metagenomes. However, strain-level complexity is far from being well characterized up to date. As the indicator of strain-level complexity, metagenomic single-nucleotide polymorphisms (SNPs) have been utilized to disentangle conspecific strains. Lots of SNP-based tools have been developed to identify strains in metagenomes. However, the sufficient sequencing depth for SNP and strain-level analysis remains unclear. We conducted ultra-deep sequencing of the human gut microbiome and constructed an unbiased framework to perform reliable SNP analysis. SNP profiles of the human gut metagenome by ultra-deep sequencing were obtained. SNPs identified from conventional and ultra-deep sequencing data were thoroughly compared and the relationship between SNP identification and sequencing depth were investigated. The results show that the commonly used shallow-depth sequencing is incapable to support a systematic metagenomic SNP discovery. In contrast, ultra-deep sequencing could detect more functionally important SNPs, which leads to reliable downstream analyses and novel discoveries. We also constructed a machine learning model to provide guidance for researchers to determine the optimal sequencing depth for their projects (SNPsnp, https://github.com/labomics/SNPsnp). To conclude, the SNP profiles based on ultra-deep sequencing data extend current knowledge on metagenomics and highlights the importance of evaluating sequencing depth before starting SNP analysis. This study provides new ideas and references for future strain-level investigations.

MetaLogo: a heterogeneity-aware sequence logo generator and aligner.

Sequence logos are used to visually display conservations and variations in short sequences. They can indicate the fixed patterns or conserved motifs in a batch of DNA or protein sequences. However, most of the popular sequence logo generators are based on the assumption that all the input sequences are from the same homologous group, which will lead to an overlook of the heterogeneity among the sequences during the sequence logo making process. Heterogeneous groups of sequences may represent clades of different evolutionary origins, or genes families with different functions. Therefore, it is essential to divide the sequences into different phylogenetic or functional groups to reveal their specific sequence motifs and conservation patterns. To solve these problems, we developed MetaLogo, which can automatically cluster the input sequences after multiple sequence alignment and phylogenetic tree construction, and then output sequence logos for multiple groups and aligned them in one figure. User-defined grouping is also supported by MetaLogo to allow users to investigate functional motifs in a more delicate and dynamic perspective. MetaLogo can highlight both the homologous and nonhomologous sites among sequences. MetaLogo can also be used to annotate the evolutionary positions and gene functions of unknown sequences, together with their local sequence characteristics. We provide users a public MetaLogo web server (http://metalogo.omicsnet.org), a standalone Python package (https://github.com/labomics/MetaLogo), and also a built-in web server available for local deployment. Using MetaLogo, users can draw informative, customized and publishable sequence logos without any programming experience to present and investigate new knowledge on specific sequence sets.

Comprehensive Strain-Level Analysis of the Gut Microbe Faecalibacterium prausnitzii in Patients with Liver Cirrhosis.

Liver cirrhosis (LC) has been associated with gut microbes. However, the strain diversity of species and its association with LC have received little attention. Here, we constructed a computational framework to study the strain heterogeneity in the gut microbiome of patients with LC. Only Faecalibacterium prausnitzii shows different single-nucleotide polymorphism (SNP) patterns between the LC and healthy control (HC) groups. Strain diversity analysis discovered that although most F. prausnitzii genomes are more deficient in the LC group than in the HC group at the strain level, a subgroup of 19 F. prausnitzii strains showed no sensitivity to LC, which is inconsistent with the species-level result. The functional differences between this subgroup and other strains may involve short-chain fatty acid production and chlorine-related pathways. These findings demonstrate functional differences among F. prausnitzii subgroups, which extend current knowledge about strain heterogeneity and relationships between F. prausnitzii and LC at the strain level. IMPORTANCE Most metagenomic studies focus on microbes at the species level, thus ignoring the different effects of different strains of the same species on the host. In this study, we explored the different microbes at the strain level in the intestines of patients with liver cirrhosis and of healthy people. Previous studies have shown that the species Faecalibacterium prausnitzii has a lower abundance in patients with liver cirrhosis than in healthy people. However, our results found multiple F. prausnitzii strains that do not decrease in abundance in patients with liver cirrhosis. It is more sensitive to select the appropriate strains as indicators to distinguish between the disease and the control samples than to use the entire species as an indicator. We clustered multiple F. prausnitzii strains and discuss the functional differences of different clusters. Our findings suggest that more attention should be paid to metagenomic studies at the strain level.

Dissecting transcriptional heterogeneity in primary gastric adenocarcinoma by single cell RNA sequencing.

OBJECTIVE: Tumour heterogeneity represents a major obstacle to accurate diagnosis and treatment in gastric adenocarcinoma (GA). Here, we report a systematic transcriptional atlas to delineate molecular and cellular heterogeneity in GA using single-cell RNA sequencing (scRNA-seq). DESIGN: We performed unbiased transcriptome-wide scRNA-seq analysis on 27 677 cells from 9 tumour and 3 non-tumour samples. Analysis results were validated using large-scale histological assays and bulk transcriptomic datasets. RESULTS: Our integrative analysis of tumour cells identified five cell subgroups with distinct expression profiles. A panel of differentiation-related genes reveals a high diversity of differentiation degrees within and between tumours. Low differentiation degrees can predict poor prognosis in GA. Among them, three subgroups exhibited different differentiation grade which corresponded well to histopathological features of Lauren's subtypes. Interestingly, the other two subgroups displayed unique transcriptome features. One subgroup expressing chief-cell markers (eg, LIPF and PGC) and RNF43 with Wnt/ß-catenin signalling pathway activated is consistent with the previously described entity fundic gland-type GA (chief cell-predominant, GA-FG-CCP). We further confirmed the presence of GA-FG-CCP in two public bulk datasets using transcriptomic profiles and histological images. The other subgroup specifically expressed immune-related signature genes (eg, LY6K and major histocompatibility complex class II) with the infection of Epstein-Barr virus. In addition, we also analysed non-malignant epithelium and provided molecular evidences for potential transition from gastric chief cells into MUC6+TFF2+ spasmolytic polypeptide expressing metaplasia. CONCLUSION: Altogether, our study offers valuable resource for deciphering gastric tumour heterogeneity, which will provide assistance for precision diagnosis and prognosis.

Susceptibility Factors of Stomach for SARS-CoV-2 and Treatment Implication of Mucosal Protective Agent in COVID-19.

Objectives: This work aims to study the gastrointestinal (GI) symptoms in severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2)-infected patients and the susceptibility factors of the stomach for SARS-CoV-2. Materials and Methods: We investigated the SARS-CoV-2 susceptibility by analyzing the expression distribution of viral entry-associated genes, ACE2 and TMPRSS2, in single-cell RNA sequencing data derived from 12 gastric mucosa samples. We also analyzed the epidemiological, demographic, clinical, and laboratory data of 420 cases with SARS-CoV-2-caused coronavirus disease 2019 (COVID-19). Results: ACE2 and TMPRSS2 are specifically expressed in enterocytes which are mainly from gastric mucosa samples with Helicobacter pylori (H. pylori) infection history and intestinal metaplasia (IM). A total of 420 patients were surveyed, of which 62 were with and 358 were without GI symptoms. There is a significant difference in average hospital stay (p < 0.001) and cost (p < 0.001) between the two groups. Among 23 hospitalized patients including seven with upper GI symptoms and 16 with lower GI symptoms, six (85.7%) and five (31.3%) had H. pylori infection history, respectively (p = 0.03). Of 18 hospitalized patients with initial upper GI symptoms, none of the eight patients with mucosal protective agent therapy (e.g., sucralfate suspension gel, hydrotalcite tablets) had diarrhea subsequently, whereas six out of 10 patients without mucosal protective agent therapy had diarrhea subsequently (p = 0.01). Conclusion: IM and H. pylori infection history may be susceptibility factors of SARS-CoV-2, and the mucosal protective agent may be useful for the blockade of SARS-CoV-2 transmission from the stomach to the intestine.

Predicting the regrowth of clinically non-functioning pituitary adenoma with a statistical model.

BACKGROUND: Compared with clinically functioning pituitary adenoma (FPA), clinically non-functioning pituitary adenoma (NFPA) lacks of detectable hypersecreting serum hormones and related symptoms which make it difficult to predict the prognosis and monitoring for postoperative tumour regrowth. We aim to investigate whether the expression of selected tumour-related proteins and clinical features could be used as tumour markers to effectively predict the regrowth of NFPA. METHOD: Tumour samples were collected from 295 patients with NFPA from Beijing Tiantan Hospital. The expression levels of 41 tumour-associated proteins were assessed using tissue microarray analyses. Clinical characteristics were analysed via univariate and multivariate logistic regression analyses. Logistic regression algorithm was applied to build a prediction model based on the expression levels of selected proteins and clinical signatures, which was then assessed in the testing set. RESULTS: Three proteins and two clinical signatures were confirmed to be significantly related to the regrowth of NFPA, including cyclin-dependent kinase inhibitor 2A (CDKN2A/p16), WNT inhibitory factor 1 (WIF1), tumour growth factor beta (TGF-ß), age and tumour volume. A prediction model was generated on the training set, which achieved a fivefold predictive accuracy of 81.2%. The prediction ability was validated on the testing set with an accuracy of 83.9%. The area under the receiver operating characteristic curves (AUC) for the signatures were 0.895 and 0.881 in the training and testing sets, respectively. CONCLUSION: The prediction model could effectively predict the regrowth of NFPA, which may facilitate the prognostic evaluation and guide early interventions.

Transcriptome Alterations in Liver Metastases of Colorectal Cancer After Acquired Resistance to Cetuximab.

BACKGROUND/AIM: Cetuximab in combination with chemotherapy is recommended as first-line therapy for metastatic colorectal cancer (mCRC) with wild-type RAS. However, drug resistance to cetuximab exists widely in mCRC and reduces the prognosis of patients. Although some genomic alterations have been demonstrated to drive acquired resistance to cetuximab, the overall compendium of inherent molecular mechanisms is still incomplete. MATERIALS AND METHODS: Four liver metastasis biopsies were collected from two mCRC patients who were treated with cetuximab in combination with 5-fluororacil plus leucovorin and oxaliplatin (FOLFOX) regimen. RESULTS: Transcriptomic analysis revealed global gene expression alterations between paired samples prior to treatment and after acquired resistance. Further bioinformatics analysis discovered differentially expressed protein-coding genes/lncRNAs/miRNAs, potential miRNA-mRNA regulatory networks and lncRNA-mRNA competing endogenous RNA network, which may be potential biomarkers or play roles during the process of acquired resistance to cetuximab. CONCLUSION: Our study contributes to deciphering the molecular mechanisms of acquired resistance to cetuximab.

Up-regulation of FAM64A promotes epithelial-to-mesenchymal transition and enhances stemness features in breast cancer cells.

FAM64A was found to be markedly up-regulated in tumor samples and associated with worse overall survival in multiple cancer types, including breast cancer. However, the functional significance of FAM64A in breast cancer remains largely unknown. In this study, we systematically investigated the expression of FAM64A in multiple public breast cancer datasets. We found that FAM64A is significantly positively correlated with tumor stemness index in breast cancer samples, corresponding with an advanced clinical grade, metastasis and unfavorable prognosis. In vitro experiments further showed an up-regulation of stemness genes after over-expressing FAM64A in breast cancer cells. FAM64A overexpression also promoted breast cancer cell proliferation, migration, accompanied by the activation of epithelial-to-mesenchymal transition (EMT). Besides, we identified a strong association of FAM64A expression with TP53 mutations in TCGA and three additional breast cancer datasets. In summary, our study revealed a novel function of FAM64A in promoting breast cancer stemness and EMT, suggesting that targeting of FAM64A may have therapeutic values in advanced breast cancer.

Dynamic testing of stimulative and suppressive biomarkers on peripheral blood cells at early stages of immunotherapy predicts response in advanced cancer patients.

PURPOSE: Immunotherapy against malignant tumors has shown considerable clinical efficacy, especially agents targeting the programmed death 1 (PD-1) pathway. In this study, we set out to dynamically determine the relationships between clinical benefit and changes in immune biomarkers on peripheral blood monocular cells (PBMCs) and try to establish a model to predict the response at an early stage of treatment. EXPERIMENTAL DESIGN: All patients recruited from December 2016 to July 2017 were treated in the Cancer Center of the Chinese People's Liberation Army General Hospital with nivolumab or pembrolizumab, and with or without chemotherapy. We investigated nine checkpoint molecules PD-1, CTLA-4, TIM-3, LAG-3, BTLA, CD160, Ki-67, OX40, and GITR on CD4+, CD8+, and NK cells by flow cytometry in patients before and after receiving each treatment of anti-PD-1 agents. RESULTS: We found that the responder group showed higher expressions of PD-1 on CD4+ and NK cells than the non-responder group after the first cycle of immunotherapy, and lower expression of CTLA-4, GITR, and OX40 after the second cycle of immunotherapy. A simple model of the combination of biomarkers was generated to predict the immunotherapeutic effect, revealing that elevation of key biomarkers after the first cycle of immunotherapy, followed by a decrease in their expression after the second cycle, was associated with a better outcome from immunotherapy at an early stage of treatment of cancer. CONCLUSION: Our work indicates that by testing biomarkers on patients' PBMCs at an early stage of treatment, the immunotherapeutic effect can be predicted, thus improving patient outcomes and cost efficiency.

Expression of Bcl-2 is a favorable prognostic biomarker in lung squamous cell carcinoma.

Lung squamous cell carcinoma (LUSC) is the second major type of lung cancer globally. The majority of patients with LUSC are clinically diagnosed at the advanced stages, thus it is urgent to identify suitable prognostic markers for LUSC. B-cell lymphoma 2 (Bcl-2) has been widely studied in non-small cell lung cancer (NSCLC). However, the prognostic role of Bcl-2 in NSCLC remains conflicting and controversial, particularly for LUSC. Although certain studies have been performed to identify the prognostic value of Bcl-2, to the best of our knowledge, no study has investigated the prognostic role of Bcl-2 in LUSC specifically. The present study aimed to comprehensively evaluate the prognostic value of Bcl-2 in LUSC. Microarray data for LUSC were downloaded from public databases, including the Gene Expression Omnibus and The Cancer Genome Atlas. Microarray data of 901 patients with LUSC from 16 data sets were retrieved. The meta-z algorithm was applied and the combined z score was identified as -2.43, suggesting Bcl-2 is a favorable prognostic biomarker. Furthermore, immunohistochemical staining of Bcl-2 expression was performed in a tissue microarray of 72 patients with LUSC and survival analysis demonstrated that patients with high expression Bcl-2 exhibited significantly more improved overall survival rates compared with those with low Bcl-2 expression. Multivariate Cox regression revealed that high expression of Bcl-2 is an independent favorable prognostic factor (hazard ratio, 0.295; confidence interval, 0.097-0.904; P<0.05). Therefore, the results of the present study demonstrated that Bcl-2 is a favorable prognostic biomarker in LUSC.

Signaling protein signature predicts clinical outcome of non-small-cell lung cancer.

BACKGROUND: Non-small-cell lung cancer (NSCLC) is characterized by abnormalities of numerous signaling proteins that play pivotal roles in cancer development and progression. Many of these proteins have been reported to be correlated with clinical outcomes of NSCLC. However, none of them could provide adequate accuracy of prognosis prediction in clinical application. METHODS: A total of 384 resected NSCLC specimens from two hospitals in Beijing (BJ) and Chongqing (CQ) were collected. Using immunohistochemistry (IHC) staining on stored formalin-fixed paraffin-embedded (FFPE) surgical samples, we examined the expression levels of 75 critical proteins on BJ samples. Random forest algorithm (RFA) and support vector machines (SVM) computation were applied to identify protein signatures on 2/3 randomly assigned BJ samples. The identified signatures were tested on the remaining BJ samples, and were further validated with CQ independent cohort. RESULTS: A 6-protein signature for adenocarcinoma (ADC) and a 5-protein signature for squamous cell carcinoma (SCC) were identified from training sets and tested in testing sets. In independent validation with CQ cohort, patients can also be divided into high- and low-risk groups with significantly different median overall survivals by Kaplan-Meier analysis, both in ADC (31 months vs. 87 months, HR 2.81; P <  0.001) and SCC patients (27 months vs. not reached, HR 9.97; P <  0.001). Cox regression analysis showed that both signatures are independent prognostic indicators and outperformed TNM staging (ADC: adjusted HR 3.07 vs. 2.43, SCC: adjusted HR 7.84 vs. 2.24). Particularly, we found that only the ADC patients in high-risk group significantly benefited from adjuvant chemotherapy (P = 0.018). CONCLUSIONS: Both ADC and SCC protein signatures could effectively stratify the prognosis of NSCLC patients, and may support patient selection for adjuvant chemotherapy.

Corrigendum: BCIP: A gene-centered platform for identifying potential regulatory genes in breast cancer.

This corrects the article DOI: 10.1038/srep45235.

Combining multi-dimensional data to identify key genes and pathways in gastric cancer.

Gastric cancer is an aggressive cancer that is often diagnosed late. Early detection and treatment require a better understanding of the molecular pathology of the disease. The present study combined data on gene expression and regulatory levels (microRNA, methylation, copy number) with the aim of identifying key genes and pathways for gastric cancer. Data used in this study was retrieved from The Cancer Genomic Atlas. Differential analyses between gastric cancer and normal tissues were carried out using Limma. Copy number alterations were identified for tumor samples. Bimodal filtering of differentially expressed genes (DEGs) based on regulatory changes was performed to identify candidate genes. Protein-protein interaction networks for candidate genes were generated by Cytoscape software. Gene ontology and pathway analyses were performed, and disease-associated network was constructed using the Agilent literature search plugin on Cytoscape. In total, we identified 3602 DEGs, 251 differentially expressed microRNAs, 604 differential methylation-sites, and 52 copy number altered regions. Three groups of candidate genes controlled by different regulatory mechanisms were screened out. Interaction networks for candidate genes were constructed consisting of 415, 228, and 233 genes, respectively, all of which were enriched in cell cycle, P53 signaling, DNA replication, viral carcinogenesis, HTLV-1 infection, and progesterone mediated oocyte maturation pathways. Nine hub genes (SRC, KAT2B, NR3C1, CDK6, MCM2, PRKDC, BLM, CCNE1, PARK2) were identified that were presumed to be key regulators of the networks; seven of these were shown to be implicated in gastric cancer through disease-associated network construction. The genes and pathways identified in our study may play pivotal roles in gastric carcinogenesis and have clinical significance.

Transcriptional response profiles of paired tumor-normal samples offer novel perspectives in pan-cancer analysis.

Both tumor and adjacent normal tissues are valuable in cancer research. Transcriptional response profiles represent the changes of gene expression levels between paired tumor and adjacent normal tissues. In this study, we performed a pan-cancer analysis based on the transcriptional response profiles from 633 samples across 13 cancer types. We obtained two interesting results. Using consensus clustering method, we characterized ten clusters with distinct transcriptional response patterns and enriched pathways. Notably, head and neck squamous cell carcinoma was divided in two subtypes, enriched in cell cycle-related pathways and cell adhesion-related pathways respectively. The other interesting result is that we identified 92 potential pan-cancer genes that were consistently upregulated across multiple cancer types. Knockdown of FAM64A or TROAP inhibited the growth of cancer cells, suggesting that these genes may promote tumor development and are worthy of further validations. Our results suggest that transcriptional response profiles of paired tumor-normal tissues can provide novel perspectives in pan-cancer analysis.

BCIP: a gene-centered platform for identifying potential regulatory genes in breast cancer.

Breast cancer is a disease with high heterogeneity. Many issues on tumorigenesis and progression are still elusive. It is critical to identify genes that play important roles in the progression of tumors, especially for tumors with poor prognosis such as basal-like breast cancer and tumors in very young women. To facilitate the identification of potential regulatory or driver genes, we present the Breast Cancer Integrative Platform (BCIP, http://omics.bmi.ac.cn/bcancer/). BCIP maintains multi-omics data selected with strict quality control and processed with uniform normalization methods, including gene expression profiles from 9,005 tumor and 376 normal tissue samples, copy number variation information from 3,035 tumor samples, microRNA-target interactions, co-expressed genes, KEGG pathways, and mammary tissue-specific gene functional networks. This platform provides a user-friendly interface integrating comprehensive and flexible analysis tools on differential gene expression, copy number variation, and survival analysis. The prominent characteristic of BCIP is that users can perform analysis by customizing subgroups with single or combined clinical features, including subtypes, histological grades, pathologic stages, metastasis status, lymph node status, ER/PR/HER2 status, TP53 mutation status, menopause status, age, tumor size, therapy responses, and prognosis. BCIP will help to identify regulatory or driver genes and candidate biomarkers for further research in breast cancer.

Gut metagenomes of type 2 diabetic patients have characteristic single-nucleotide polymorphism distribution in Bacteroides coprocola.

BACKGROUND: Gut microbes play a critical role in human health and disease, and researchers have begun to characterize their genomes, the so-called gut metagenome. Thus far, metagenomics studies have focused on genus- or species-level composition and microbial gene sets, while strain-level composition and single-nucleotide polymorphism (SNP) have been overlooked. The gut metagenomes of type 2 diabetes (T2D) patients have been found to be enriched with butyrate-producing bacteria and sulfate reduction functions. However, it is not known whether the gut metagenomes of T2D patients have characteristic strain patterns or SNP distributions. FINDINGS: We downloaded public gut metagenome datasets from 170 T2D patients and 174 healthy controls and performed a systematic comparative analysis of their metagenome SNPs. We found that Bacteroides coprocola, whose relative abundance did not differ between the groups, had a characteristic distribution of SNPs in the T2D patient group. We identified 65 genes, all in B. coprocola, that had remarkably different enrichment of SNPs. The first and sixth ranked genes encode glycosyl hydrolases (GenBank accession EDU99824.1 and EDV02301.1). Interestingly, alpha-glucosidase, which is also a glycosyl hydrolase located in the intestine, is an important drug target of T2D. These results suggest that different strains of B. coprocola may have different roles in human gut and a specific set of B. coprocola strains are correlated with T2D.

miR-21-mediated Radioresistance Occurs via Promoting Repair of DNA Double Strand Breaks.

miR-21, as an oncogene that overexpresses in most human tumors, is involved in radioresistance; however, the mechanism remains unclear. Here, we demonstrate that miR-21-mediated radioresistance occurs through promoting repair of DNA double strand breaks, which includes facilitating both non-homologous end-joining (NHEJ) and homologous recombination repair (HRR). The miR-21-promoted NHEJ occurs through targeting GSK3B (a novel target of miR-21), which affects the CRY2/PP5 pathway and in turn increases DNA-PKcs activity. The miR-21-promoted HRR occurs through targeting both GSK3B and CDC25A (a known target of miR-21), which neutralizes the effects of targeting GSK3B-induced CDC25A increase because GSK3B promotes degradation of both CDC25A and cyclin D1, but CDC25A and cyclin D1 have an opposite effect on HRR. A negative correlation of expression levels between miR-21 and GSK3ß exists in a subset of human tumors. Our results not only elucidate miR-21-mediated radioresistance, but also provide potential new targets for improving radiotherapy.