Applications of Rna Sequencing

Large-scale biomedical datasets that combine histology with paired RNA sequencing (RNA-seq) and Whole Genome Sequencing (WGS) data creates the opportunity to understand ways that somatic mutations and variation in gene expression influence tissue-level properties in health and disease. A new study led by Francesco Cisternino details the development of a Vision Transformer trained on 1.7 million histology images across 23 healthy tissue types that can be used for automatic tissue segmentation and the prediction of spatially localized RNA expression levels from H&E histology images. The authors show that by learning self-supervised representations from a large set of histology images across healthy tissues, the model can automatically identify tissue substructures and pathologies without labels. Self-supervised learning for characterising histomorphological diversity and spatial RNA expression prediction across 23 human tissue types. https://lnkd.in/e6i_sb8S Methods overview: The authors used 13,898 whole slide images from 23 tissue types in 838 donors from the GTEx project. They preprocessed the images by segmenting the tissue from background using a U-net, and tiled the tissue into 63 x 63 μm2 regions. They trained a small Vision Transformer (ViT-S) using the self-supervised DINO framework on 1.7 million GTEx histology tiles to extract morphological features.  Using the learned features, they classified tiles through a K-Nearest Neighbors model to derive phenotypes in terms of extent of detected regions. The RNAPath model takes tile embeddings as input and predicts both local (tile-level) and global (sample-level) gene expression as output, along with a heatmap to visualize predicted spatial gene activity. Results overview: The self-supervised DINO embeddings show better qualitative clustering and 43% improvement in silhouette score compared to other representation learning methods. Using a kNN approach, they are able to automatically segment whole slide images into constituent tissue substructures and pathology proportions with high accuracy. They find substantial variability in tissue substructure proportions across donors within the same tissue type. Using the substructure and pathology proportions, they identify profound tissue variability across donors that drives substantial differential gene expression, as well as characterize germline genetic variants associated with specific histopathological features. The RNAPath models are able to predict individual RNA expression levels from histology with superior performance to competing methods. They validate RNAPath spatial predictions using positive control immunohistochemistry and characterize the localized expression signatures of 29 individual substructures and pathologies. Both the histology tile representations and RNAPath generalize well to an external validation cohort, TCGA-BRCA, demonstrating ability to segment carcinoma from benign tissue.

Bioinformatics gold is often found in the junk pile. Here’s how to mine it. 1/ You delete them without thinking. Unmapped reads. But buried in the trash... is treasure. Let me explain. 2/ Unmapped RNA-seq reads aren’t useless. They often hold microbial RNA. Think: hidden infections, gut bugs, or viral contaminants. 3/ You can recover microbial reads from human RNAseq That’s data most people throw away. Don’t be most people. 4/ Or take circular RNAs (circRNAs). Standard aligners can’t map back-spliced reads. But those “unmapped” reads? They are the circRNAs. 5/ circRNAs regulate gene expression and are potential biomarkers for cancer, brain disorders, and more. 6/ reads can not be mapped to the highly variable region of TCR and BCR genes, but you can use tools such as https://lnkd.in/e-5jT5MS to reconstruct the TCR and BCR sequences using the unmapped reads. Those are valuable data for studying the immune response. I used it! 7/ RNA-seq is more than gene counts. You can call SNPs and somatic mutations. Even BRAF V600E shows up if you know where to look. 8/ You don’t always need DNA-seq to call variants. RNA-seq gives you heterozygous SNPs (like rsIDs) in expressed regions—cheaply. I identified sample swaps for multi-omics studies with both RNAseq and WES data by calling SNPs from RNAseq data and mapping them to the WES SNPs. 9/ You can even phase reads and detect allele-specific expression (ASE). Want to study imprinting? IGF2, H19? Start here. 10/ ASE tells you which allele is active—maternal or paternal. It reveals silencing, imprinting, or regulatory variation. That’s power. 11/ Whole-exome sequencing (WES)? not only for detecting mutations. Use coverage data to detect copy number variants (CNVs). 12/ ATAC-seq isn’t just for open chromatin. You can extract CNV patterns from it, especially in tumors. or detect ecDNA too! 13/ Even repeats have meaning. Unmapped RNA-seq reads often come from LINE/SINE retrotransposons. Like Alu or L1. 14/ These elements are noisy—but informative. Dysregulated transposons are linked to neurodegeneration, cancer, aging. 15/ You can even use RNAseq data to determine variable 3UTR length and that has implications in cancer too! 14/ Key takeaways: Unmapped ≠ useless RNA-seq is a multi-tool: SNPs, circRNA, microbes, ASE, 3UTR length WES & ATAC-seq reveal CNVs Repeats matter Bioinformatics rewards the curious. Even in the trash. I hope you've found this post helpful. Follow me for more. Subscribe to my FREE newsletter chatomics to learn bioinformatics https://lnkd.in/erw83Svn

We might be looking at brain cancer all wrong.  Most therapies target genetic mutations, but what if the key to overcoming treatment resistance lies in how tumor cells shift identities? Cell states, not just mutations, may be the missing piece in our fight against brain cancer. I just published my thoughts on why transcriptional signatures could revolutionize brain cancer treatment. Recent research in Nature Precision Oncology reveals glioma cells exist in a three-dimensional landscape where they adapt to current treatments by changing their cellular states. Many believe we need expensive single-cell sequencing to detect these shifts - but that's not the case. Standard bulk RNA sequencing (already being generated at scale by companies like Tempus) can identify these signatures in clinical practice, making this approach immediately actionable. This creates an entirely new paradigm: instead of just targeting what a tumor is today, we can block what it's trying to become tomorrow. But there's one critical challenge we still need to solve... [Read more in my latest Substack post - linked in comments] Who else is working at this intersection of transcriptional signatures and precision oncology? Would love to connect with fellow data enthusiasts tackling these complex problems. s/o Célia Lemoine for the amazing work on this Nature Paper #PrecisionMedicine #CancerResearch #DataScience #Oncology