mlf-core: a framework for deterministic machine learning.

MOTIVATION: Machine learning has shown extensive growth in recent years and is now routinely applied to sensitive areas. To allow appropriate verification of predictive models before deployment, models must be deterministic. Solely fixing all random seeds is not sufficient for deterministic machine learning, as major machine learning libraries default to the usage of nondeterministic algorithms based on atomic operations. RESULTS: Various machine learning libraries released deterministic counterparts to the nondeterministic algorithms. We evaluated the effect of these algorithms on determinism and runtime. Based on these results, we formulated a set of requirements for deterministic machine learning and developed a new software solution, the mlf-core ecosystem, which aids machine learning projects to meet and keep these requirements. We applied mlf-core to develop deterministic models in various biomedical fields including a single-cell autoencoder with TensorFlow, a PyTorch-based U-Net model for liver-tumor segmentation in computed tomography scans, and a liver cancer classifier based on gene expression profiles with XGBoost. AVAILABILITY AND IMPLEMENTATION: The complete data together with the implementations of the mlf-core ecosystem and use case models are available at https://github.com/mlf-core.

Pathogen detection in RNA-seq data with Pathonoia.

BACKGROUND: Bacterial and viral infections may cause or exacerbate various human diseases and to detect microbes in tissue, one method of choice is RNA sequencing. The detection of specific microbes using RNA sequencing offers good sensitivity and specificity, but untargeted approaches suffer from high false positive rates and a lack of sensitivity for lowly abundant organisms. RESULTS: We introduce Pathonoia, an algorithm that detects viruses and bacteria in RNA sequencing data with high precision and recall. Pathonoia first applies an established k-mer based method for species identification and then aggregates this evidence over all reads in a sample. In addition, we provide an easy-to-use analysis framework that highlights potential microbe-host interactions by correlating the microbial to the host gene expression. Pathonoia outperforms state-of-the-art methods in microbial detection specificity, both on in silico and real datasets. CONCLUSION: Two case studies in human liver and brain show how Pathonoia can support novel hypotheses on microbial infection exacerbating disease. The Python package for Pathonoia sample analysis and a guided analysis Jupyter notebook for bulk RNAseq datasets are available on GitHub.

RORγt inhibitors block both IL-17 and IL-22 conferring a potential advantage over anti-IL-17 alone to treat severe asthma.

BACKGROUND: RORγt is a transcription factor that enables elaboration of Th17-associated cytokines (including IL-17 and IL-22) and is proposed as a pharmacological target for severe asthma. METHODS: IL-17 immunohistochemistry was performed in severe asthma bronchial biopsies (specificity confirmed with in situ hybridization). Primary human small airway epithelial cells in air liquid interface and primary bronchial smooth muscle cells were stimulated with recombinant human IL-17 and/or IL-22 and pro-inflammatory cytokines measured. Balb/c mice were challenged intratracheally with IL-17 and/or IL-22 and airway hyperreactivity, pro-inflammatory cytokines and airway neutrophilia measured. Balb/c mice were sensitized intraperitoneally and challenged intratracheally with house dust mite extract and the effect of either a RORγt inhibitor (BIX119) or an anti-IL-11 antibody assessed on airway hyperreactivity, pro-inflammatory cytokines and airway neutrophilia measured. RESULTS: We confirmed in severe asthma bronchial biopsies both the presence of IL-17-positive lymphocytes and that an IL-17 transcriptome profile in a severe asthma patient sub-population. Both IL-17 and IL-22 stimulated the release of pro-inflammatory cytokine and chemokine release from primary human lung cells and in mice. Furthermore, IL-22 in combination with IL-17, but neither alone, elicits airway hyperresponsiveness (AHR) in naïve mice. A RORγt inhibitor specifically blocked both IL-17 and IL-22, AHR and neutrophilia in a mouse house dust mite model unlike other registered or advanced pipeline modes of action. Full efficacy versus these parameters was associated with 90% inhibition of IL-17 and 50% inhibition of IL-22. In contrast, anti-IL-17 also blocked IL-17, but not IL-22, AHR or neutrophilia. Moreover, the deregulated genes in the lungs from these mice correlated well with deregulated genes from severe asthma biopsies suggesting that this model recapitulates significant severe asthma-relevant biology. Furthermore, these genes were reversed upon RORγt inhibition in the HDM model. Cell deconvolution suggested that the responsible cells were corticosteroid insensitive γδ-T-cells. CONCLUSION: These data strongly suggest that both IL-17 and IL-22 are required for Th2-low endotype associated biology and that a RORγt inhibitor may provide improved clinical benefit in a severe asthma sub-population of patients by blocking both IL-17 and IL-22 biology compared with blocking IL-17 alone.

A Multi-omics Data Resource for Frontotemporal Dementia Research.

Frontotemporal dementia (FTD) is a neurodegenerative disease with high heritability. Almost half of all familial cases are caused by mutations in one of the three genes MAPT, GRN and C9orf72. Even though major advances in FTD research have been achieved during the last decades, it is not yet fully understood how mutations in these diverse genes lead to the disease. To improve our understanding of FTD, the Risk and Modifying Factors in Frontotemporal Dementia (RiMod-FTD) consortium has created an FTD-specific multi-omics data resource. Using multiple omics technologies on post-mortem brain tissue from patients with mutations in GRN, MAPT or C9orf72 and healthy controls, the resource aims to provide a comprehensive cellular profile of FTD. Furthermore, brain tissue from multiple mouse models and induced pluripotent stem cells (iPSC)-derived neuronal cultures were profiled with similar multi-omics technologies to make up for the shortcomings of post-mortem brain tissue. All data are publicly available to all researchers, and ongoing efforts aim to increase the available datasets and to improve their accessibility. The RiMod-FTD resource represents a uniquely valuable dataset for the field of FTD research, which we hope will accelerate the scientific progress in the field.

First insights into the nature and evolution of antisense transcription in nematodes.

BACKGROUND: The development of multicellular organisms is coordinated by various gene regulatory mechanisms that ensure correct spatio-temporal patterns of gene expression. Recently, the role of antisense transcription in gene regulation has moved into focus of research. To characterize genome-wide patterns of antisense transcription and to study their evolutionary conservation, we sequenced a strand-specific RNA-seq library of the nematode Pristionchus pacificus. RESULTS: We identified 1112 antisense configurations of which the largest group represents 465 antisense transcripts (ASTs) that are fully embedded in introns of their host genes. We find that most ASTs show homology to protein-coding genes and are overrepresented in proteomic data. Together with the finding, that expression levels of ASTs and host genes are uncorrelated, this indicates that most ASTs in P. pacificus do not represent non-coding RNAs and do not exhibit regulatory functions on their host genes. We studied the evolution of antisense gene pairs across 20 nematode genomes, showing that the majority of pairs is lineage-specific and even the highly conserved vps-4, ddx-27, and sel-2 loci show abundant structural changes including duplications, deletions, intron gains and loss of antisense transcription. In contrast, host genes in general, are remarkably conserved and encode exceptionally long introns leading to unusually large blocks of conserved synteny. CONCLUSIONS: Our study has shown that in P. pacificus antisense transcription as such does not define non-coding RNAs but is rather a feature of highly conserved genes with long introns. We hypothesize that the presence of regulatory elements imposes evolutionary constraint on the intron length, but simultaneously, their large size makes them a likely target for translocation of genomic elements including protein-coding genes that eventually end up as ASTs.

A multi-omics dataset for the analysis of frontotemporal dementia genetic subtypes.

Understanding the molecular mechanisms underlying frontotemporal dementia (FTD) is essential for the development of successful therapies. Systematic studies on human post-mortem brain tissue of patients with genetic subtypes of FTD are currently lacking. The Risk and Modyfing Factors of Frontotemporal Dementia (RiMod-FTD) consortium therefore has generated a multi-omics dataset for genetic subtypes of FTD to identify common and distinct molecular mechanisms disturbed in disease. Here, we present multi-omics datasets generated from the frontal lobe of post-mortem human brain tissue from patients with mutations in MAPT, GRN and C9orf72 and healthy controls. This data resource consists of four datasets generated with different technologies to capture the transcriptome by RNA-seq, small RNA-seq, CAGE-seq, and methylation profiling. We show concrete examples on how to use the resulting data and confirm current knowledge about FTD and identify new processes for further investigation. This extensive multi-omics dataset holds great value to reveal new research avenues for this devastating disease.

Distinct cell type-specific protein signatures in GRN and MAPT genetic subtypes of frontotemporal dementia.

Frontotemporal dementia is characterized by progressive atrophy of frontal and/or temporal cortices at an early age of onset. The disorder shows considerable clinical, pathological, and genetic heterogeneity. Here we investigated the proteomic signatures of frontal and temporal cortex from brains with frontotemporal dementia due to GRN and MAPT mutations to identify the key cell types and molecular pathways in their pathophysiology. We compared patients with mutations in the GRN gene (n = 9) or with mutations in the MAPT gene (n = 13) with non-demented controls (n = 11). Using quantitative proteomic analysis on laser-dissected tissues we identified brain region-specific protein signatures for both genetic subtypes. Using published single cell RNA expression data resources we deduced the involvement of major brain cell types in driving these different protein signatures. Subsequent gene ontology analysis identified distinct genetic subtype- and cell type-specific biological processes. For the GRN subtype, we observed a distinct role for immune processes related to endothelial cells and for mitochondrial dysregulation in neurons. For the MAPT subtype, we observed distinct involvement of dysregulated RNA processing, oligodendrocyte dysfunction, and axonal impairments. Comparison with an in-house protein signature of Alzheimer's disease brains indicated that the observed alterations in RNA processing and oligodendrocyte function are distinct for the frontotemporal dementia MAPT subtype. Taken together, our results indicate the involvement of different brain cell types and biological mechanisms in genetic subtypes of frontotemporal dementia. Furthermore, we demonstrate that comparison of proteomic profiles of different disease entities can separate general neurodegenerative processes from disease-specific pathways, which may aid the development of disease subtype-specific treatment strategies.

Evidence for GRN connecting multiple neurodegenerative diseases.

Previous research using genome-wide association studies has identified variants that may contribute to lifetime risk of multiple neurodegenerative diseases. However, whether there are common mechanisms that link neurodegenerative diseases is uncertain. Here, we focus on one gene, GRN, encoding progranulin, and the potential mechanistic interplay between genetic risk, gene expression in the brain and inflammation across multiple common neurodegenerative diseases. We utilized genome-wide association studies, expression quantitative trait locus mapping and Bayesian colocalization analyses to evaluate potential causal and mechanistic inferences. We integrate various molecular data types from public resources to infer disease connectivity and shared mechanisms using a data-driven process. Expression quantitative trait locus analyses combined with genome-wide association studies identified significant functional associations between increasing genetic risk in the GRN region and decreased expression of the gene in Parkinson's, Alzheimer's and amyotrophic lateral sclerosis. Additionally, colocalization analyses show a connection between blood-based inflammatory biomarkers relating to platelets and GRN expression in the frontal cortex. GRN expression mediates neuroinflammation function related to multiple neurodegenerative diseases. This analysis suggests shared mechanisms for Parkinson's, Alzheimer's and amyotrophic lateral sclerosis.

Deep learning-based cell composition analysis from tissue expression profiles.

We present Scaden, a deep neural network for cell deconvolution that uses gene expression information to infer the cellular composition of tissues. Scaden is trained on single-cell RNA sequencing (RNA-seq) data to engineer discriminative features that confer robustness to bias and noise, making complex data preprocessing and feature selection unnecessary. We demonstrate that Scaden outperforms existing deconvolution algorithms in both precision and robustness. A single trained network reliably deconvolves bulk RNA-seq and microarray, human and mouse tissue expression data and leverages the combined information of multiple datasets. Because of this stability and flexibility, we surmise that deep learning will become an algorithmic mainstay for cell deconvolution of various data types. Scaden's software package and web application are easy to use on new as well as diverse existing expression datasets available in public resources, deepening the molecular and cellular understanding of developmental and disease processes.