Search bioRxiv⌕ Search
Home/Archive/bioRxiv

bioRxiv · 10.1101/2022.06.07.495072

Trans-Driver: a deep learning approach for cancer driver gene discovery with multi-omics data

Yang, H.·Zhang, L.·Zhou, D.·Li, D.·Zhang, J.·Wang, Z.
View original source ↗Export metadata ↓

Abstract

Driver genes play a crucial role in the growth of cancer cells. Accurate identification of cancer driver genes is helping to strengthen the understanding of cancer pathogenesis and is conducive to the development of cancer treatment and drug-targe driver genes. However, due to the diversity and complexity of the multi-omics data, it is still challenging to identify cancer drivers.In this study, we propose Trans-Driver, a deep supervised learning method with a novel transformer network, which integrates multi-omics data to learn the differences and associations between different omics data for cancer drivers discovery. Compared with other state-of-the-art driver gene identification methods, Trans-Driver has achieved excellent performance on TCGA and CGC data Machine learning for multi-omics data integration in cancer. Among 20,000 protein-coding genes, Trans-Driver reported 185 candidate driver genes, of which 103 genes (about 55%) were included in the gold standard CGC data set. Finally, we analyzed the contribution of each feature to the identification of driver genes. We found that the integration of multi-omics data can improve the performance of our method compared with using only somatic mutation data. Through detailed analysis, we found that the candidate drivers are clinically meaningful, proving the practicability of Trans-Driver. Author summaryMany methods have been developed to identify cancer driver genes. However, most of these methods use single-omics data for cancer driver gene identification. Multi-omics-based methods for cancer driver gene identification are rare. Trans-Driver uses deep learning to process multi-omics data and learn the relationships between multi-omics data for cancer driver gene prediction. We have predicted 185 candidate cancer driver genes out of among 20,000 protein-coding genes. Also, we performed cancer driver gene prediction on 33 cancer types, and we identified the cancer driver genes corresponding to each cancer type. And, we observed that the predicted cancer driver genes were shown to have a role in cancer progression in recent studies. Our proposed method for cancer driver gene identification using multi-omics data has improved performance compared to using mutation data alone.

Source connections

Explore related subjects

bioinformaticsPreprint

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Yang, H., Zhang, L., Zhou, D., Li, D., Zhang, J., Wang, Z.. 2022-06-09. Trans-Driver: a deep learning approach for cancer driver gene discovery with multi-omics data. https://doi.org/10.1101/2022.06.07.495072

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

Algorithms for Designing Therapeutic Oligonucleotides on Pangenome Graphs

Designing therapeutic oligonucleotides that act reliably across human genetic diversity requires more than alignment to a single reference genome. Binding affinity of an siRNA or ASO to a usually complementary site in an mRNA target molecule is one main determinant of drug efficacy. Thus, variation in sub-populations can lead to failures in drug development if this variation reduces binding affinity. We provide a first methodological step towards principled, variant-aware design of therapeutic oligonucleotides. The genetic diversity in target genes, or specific regions thereof, is represented as a pangenome graph. We define individual or small sets of suitable therapeutic oligonucleotides by their ability to cover all paths in a directed acyclic graph within a bounded edit distance <= k. Edit distance is used as an accepted proxy to the commonly used binding affinity. We formalize the resulting computational problem and provide complexity results: NP-hardness of identifying such minimal sets even for k =0, and coNP-completeness of verifying the validity of a given set of oligonucleotides. Further, we propose algorithmic approaches operating on a product graph of an automaton (an Aho-Corasick DFA or an Ukkonen DFA) and the pangenome graph, as well as a dynamic programming-based lossless filter. We evaluate the approaches on gene-specific subgraphs of human pangenome graphs and demonstrate favorable performance compared to naive brute-force computation. Our results show that candidate target sites with a threshold of k <= 3 can be identified even for large, complex graphs where the brute-force computation did not complete within 24h. We observe that the dynamic programming-based lossless filter outperforms both product-graph approaches as well as brute-force baselines while maintaining an average approximation ratio < 1.5 across more than a hundred evaluated subgraphs. We also present examples where a small set of oligonucleotides constitutes a feasible solution where no single oligonucleotide with a threshold of k <= 2 covers all paths.

bioinformatics↗
bioRxiv2026

Evolutionarily Conserved Fragments reveal modular architecture and expand functional coverage in bacteriophage proteins

Bacteriophage proteins evolve through extensive modular exchange, yet the sequence units underlying this process remain poorly defined. Classical domains are plausible evolutionary modules, but they incompletely annotate phage proteins and may not always coincide with the regions conserved across protein families. Here, we constructed 13,035 Evolutionarily Conserved Fragments (ECFs), grouped into 3,167 families, through large-scale HMM-HMM comparisons of proteins from 5,396 RefSeq phage genomes, without imposing predefined domain boundaries. ECFs overlapped substantially with ECOD and Pfam annotations but formed a heterogeneous spectrum comprising subdomains, domain-equivalent units, composite regions and fragments with no counterpart in either resource. Their relationships were frequently nested, indicating that conserved modules occur at multiple organisational scales. ECFs annotated 50% of phage representative proteins (at least doubling the coverage of either database), with the largest gains in structural and infection-related functions, a pattern replicated in independent metagenomic data. ECF families also showed function-dependent patterns of diversity and spatial segregation and provided functional associations for 23,925 protein families lacking detectable Pfam or ECOD domains. In receptor-binding proteins, ECFs subdivided established domains, delineated previously unannotated structural elements and recovered homologous regions where structural similarity was ambiguous. These results show that the evolutionary building blocks of phage proteins do not always coincide with classical domains, highlighting that evolutionary processes act across multiple scales. ECFs provide a complementary, sequence-derived framework for investigating how complex viral proteins are assembled, diversified and functionally organised.

bioinformatics↗
bioRxiv2026

deepTools4: Faster, Scalable, and Modernized Genomic Data Analysis

The deepTools suite is widely used for quality control, summarization, and visualization of high-throughput sequencing data. However, as genomic datasets grow in scale and depth, existing tools often become computational bottlenecks. We present deepTools4, a modernized framework designed for high performance, scalability, and enhanced usability, while maintaining seamless integration with the Galaxy platform. Core functionality has been rewritten and thoroughly benchmarked, achieving substantial computational speed-ups. Visualization tools have been modernized to provide more intuitive, high-quality graphical outputs. In addition, our testing framework has been expanded to facilitate continuous development and ensure robustness. We demonstrate the utility of deepTools4 across diverse epigenomics and transcriptomics datasets, highlighting its ability to summarize genome-wide signals, apply a wide range of common normalization methods, and visualize enrichment patterns. The enhanced deepTools4 framework delivers a faster, more scalable solution for reproducible genome-wide analysis and interpretation across many sequencing modalities. Availability: https://github.com/deeptools/deepTools/

bioinformatics↗