A new study by the Krauthammer group presents DAS-Net, an explainable deep learning model for predicting disease activity in chronic inflammatory joint diseases. This model utilizes patient data to make predictions and identifies similar patients based on disease progression. The approach outperforms traditional models and enhances understanding through feature attribution, providing insights into key patient characteristics impacting disease.

The success of prime editing depends on the prime editing guide RNA (pegRNA) design and target locus. Here, we developed machine learning models that reliably predict prime editing efficiency. PRIDICT2.0 assesses the performance of pegRNAs for all edit types up to 15 bp in length in mismatch repair-deficient and mismatch repair-proficient cell lines and in vivo in primary cells. With ePRIDICT, we further developed a model that quantifies how local chrom

The Krauthammer group introduces Fragmentstein, a command-line tool designed to convert non-sensitive cfDNA fragment data into BAM files suitable for standard bioinformatics analyses. This innovative approach facilitates data sharing without compromising sensitive genomic information, enabling analyses like CNV, nucleosome occupancy, and fragment length studies while ensuring data privacy.

Tune in now on Apple Podcasts, Spotify, or Podigee.

The Krauthammer group introduces Fragmentstein, a command-line tool designed to convert non-sensitive cfDNA fragment data into BAM files suitable for standard bioinformatics analyses. This innovative approach facilitates data sharing without compromising sensitive genomic information, enabling analyses like CNV, nucleosome occupancy, and fragment length studies while ensuring data privacy.

In this work, the Polymenidou group and collaborators assessed TAR DNA-binding protein 43 (TDP-43) seeding activity and aggregates detection in olfactory mucosa of patients with frontotemporal lobar degeneration with TDP-43-immunoreactive pathology (FTLD-TDP) by TDP-43 seeding amplification assay (TDP43-SAA) and immunocytochemical analysis. They demonstrate that TDP-43 aggregates can be detectable in olfactory mucosa, suggesting that their TDP43-Seeding Amplification Assay might be useful for identifying and monitoring FTLD-TDP in living patients.

In this work, the Bodenmiller group presents an end-to-end workflow for multiplexed tissue image processing and analysis that integrates previously developed computational tools to enable image segmentation, feature extraction and spatially resolved single-cell analysis in a user-friendly and customizable way. This protocol can be implemented by researchers with basic bioinformatics training, and the analysis of the provided dataset can be completed within 5–6 h.

An extended version is available at https://bodenmillergroup.github.io/IMCDataAnalysis/.

Link to BME337: Introduction to Digital Health in the course catalogue

This course will provide an overview of tools, methods and applications driving digital health innovations. ln the course, students will gain first insights into the topic by UZH digital health experts presenting their latest research in clinical data science, imaging, mobile health, digital therapeutics, augmented reality and related domains. The course is the first one of five modules on 'Digital Heath', which aims at providing students with the fundamental knowledge about using data and computation in clinical research and digital health interventions. To prepare the students for the following modules, this introductory course will provide the necessary understanding of the key concepts in digital health at the intersection of (bio- and behavioural) medicine, data science, bioinformatics, and information systems.

Topics that will be introduced in this course include:

- General overview on how data, digital tools and Al are used in healthcare;

- Clinical data science and translational informatics;

- Biomedical lmage Analysis and Machine Learning;

- Digital Tools for Medical Knowledge and Decision Support;

- Mobile health data streams for real world evidence generation and public health;

- Digital therapeutics and health interventions for prevention, management, and treatment of diseases.

See Hosogane et al., Nature Methods

In this work, the Bodenmiller group extends Imaging mass cytometry (IMC) to low-abundance markers through incorporation of the DNA-based signal amplification by exchange reaction, immuno-SABER. Using this novel method, the tumor immune microenvironment in human melanoma was analyzed by simultaneous imaging of 18 markers with immuno-SABER and 20 markers without amplification. SABER-IMC enabled the identification of immune cell phenotypic markers that are not detectable with IMC alone.

See Strotton et al., Nature Methods

Whole-organism to atomic-level imaging is possible with tissue-penetrant, picometer-wavelength X-rays. To enable highly multiplexed X-ray imaging, the Bodenmiller group developed multielement Z-tag X-ray fluorescence (MEZ-XRF) that can operate at kHz speeds when combined with signal amplification by exchange reaction (SABER)-amplified Z-tag reagents. Parallel imaging of 20 Z-tag or SABER Z-tag reagents at subcellular resolution was demonstrated in cell lines and multiple human tissues. The unique multiscale, nondestructive nature of MEZ-XRF, combined with SABER Z-tags for high sensitivity or enhanced speed, enables highly multiplexed bioimaging across biological scales.

In clinical workflow, replacing CT with MR image enhances workflow efficiency and reduces patient radiation. To eliminate CT from the workflow,  the information provided by CT needs to be generated via an MR image. In this work, the Menze group proposes a machine learning method to generate accurate synthetic CT (sCT) from MRI with a quantitative accuracy suitable for RT dose planning application, setting the stage for a broader clinical evaluation of sCT based RT planning on different anatomical regions. 

In this work, the Bodenmiller group defined cancer-associated fibroblast (CAF) phenotypes by analysing a single-cell RNA sequencing (scRNA-seq) dataset of over 16,000 stromal cells from tumours of 14 breast cancer patients, based on which nine CAF phenotypes and one class of pericytes were defined and functionally annotated. Next, the classification system was validated in four additional cancer types and highly multiplexed imaging mass cytometry was used on matched breast cancer samples to confirm the defined CAF phenotypes at the protein level and to analyse their spatial distribution within tumours. This general CAF classification scheme will allow comparison of CAF phenotypes across studies, facilitate analysis of their functional roles, and potentially guide development of new treatment strategies in the future.

 

 

Read more about the new BioMedical Informatics Platform in UZH News: "Unlocking the Data Treasure Chest"

In this work, the Bodenmiller group uses multiplex imaging mass cytometry to analyze breast cancer patients' single tumor cell phenotypes, both in primary tumors and in matched lymph node metastases. They report high phenotypic variability between these tissue sites and identify high- and low-risk disseminated cell phenotypes and prognostic markers.
See Fischer et al., Cell Reports Medicine

In this work, the Menze group presents a multi-stage deep learning approach to predict collateral flow grading in stroke patients. Based on radiomic features extracted from MR perfusion data, a region of interest (RoI) detection task is formulated as a reinforcement learning problem and used to train a deep learning network to automatically detect the occluded region within the 3D MR perfusion volumes. Next, radiomic features are extracted from the obtained RoI through local image descriptors and denoising auto-encoders. Finally, a convolutional neural network and other machine learning classifiers are applied to the extracted radiomic features to automatically predict the collateral flow grading of the given patient volume as one of three severity classes. This automated deep learning approach is faster than visual inspection, eliminates grading bias and demonstrates a performance comparable to expert grading.

See Tetteh et al., Front Neurol

In this work, the Menze group proposes a novel Detection Transformer for 3D anatomical structure detection, dubbed Focused Decoder. Focused Decoder precisely focuses on relevant anatomical structures, using anatomical region atlas information to deploy query anchors, while at the same time restricting the cross-attention’s field of view to regions of interest. Evaluated on two publicly available CT datasets, Focused Decoder is shown to provide both strong detection results, as well as highly intuitive explainability via attention weights.

See Wittmann et al.,Journal of Machine Learning for Biomedical Imaging
Code is available at https://github.com/bwittmann/transoar

Read more about the new BioMedical Informatics Platform in UZH News: "Unlocking the Data Treasure Chest"

In this work, the Krauthammer lab and collaborators conducted a high-throughput screen to analyze prime editing outcomes of >90K pegRNAs on a set of >13K human pathogenic mutations. This dataset yielded sequence context features that influence prime editing and was subsequently used to train PRIDICT; an attention-based bidirectional recurrent neural network. PRIDICT reliably predicts editing rates for all small-sized genetic changes. Validation of PRIDICT, both on endogenous editing sites as well as on an external dataset, showed that pegRNAs with high versus low PRIDICT scores showed increased prime editing efficiencies in different cell types in vitro (12-fold) and in hepatocytes in vivo (tenfold), highlighting the value of PRIDICT for basic and translational research applications. PRIDICT is freely accessible at www.pridict.it.

See Mathis, Allam et al., Nature Biotechnology

This work is featured on UZH News: Artificial Intelligence Improves Efficiency of Genome Editing

Immune checkpoint therapy aims at preventing or reversing exhausted T cell states, but in breast cancer, T cell exhaustion is poorly understood. In this work, the Bodenmiller group used single-cell transcriptomics combined with imaging mass cytometry to systematically study immune environments of human luminal breast tumors with and without exhausted T cells. The data show that expression of PD-1 and CXCL13 on T cells, and MHC-I – but not PD-L1 – on tumor cells on tumor cells are powerful differentiators between these environments.
See Tietscher et al., Nat Communications

Hexanucleotide G₄C₂ repeat expansions in the C9orf72 gene are the most common genetic cause of amyotrophic lateral sclerosis (ALS) and frontotemporal dementia (FTD). Dipeptide repeat proteins (DPRs) generated by translation of repeat-containing RNAs are key targets for therapeutic intervention. In this work, the Polymenidou group and collaborators generated human antibodies that bind DPRs with high affinity and specificity, systematically characterized these against multiple DPR species and tested the biological effects of antibodies targeting poly-GA in different cellular and mouse models. 

See Jambeau et al, PNAS

In this work, the Menze group and collaborators report the set-up and results of the Liver Tumor Segmentation Benchmark (LiTS), which was organized in conjunction with the IEEE International Symposium on Biomedical Imaging (ISBI) 2017 and the International Conferences on Medical Image Computing and Computer-Assisted Intervention (MICCAI) 2017 and 2018. There was not a single algorithm that performed best for both liver and liver tumors in these three events. The best liver segmentation algorithm achieved a Dice score of 0.963, whereas the best algorithms for tumor segmentation achieved Dices scores of 0.674 (ISBI 2017), 0.702 (MICCAI 2017), and 0.739 (MICCAI 2018). Retrospectively, additional analysis on liver tumor detection revealed that not all top-performing segmentation algorithms worked well for tumor detection: The best liver tumor detection method achieved a lesion-wise recall of 0.458 (ISBI 2017), 0.515 (MICCAI 2017), and 0.554 (MICCAI 2018), indicating the need for further research. Nonetheless, LiTS remains an active benchmark and resource for research. See Bilic et al., Med Image Anal.
Both data and online evaluation are accessible via https://competitions.codalab.org/competitions/17094.

Multiplexed imaging methods allow simultaneous detection of dozens of proteins and hundreds of RNAs, but parameters for the design of optimal multiplex imaging studies are lacking. In this work, the Bodenmiller group developed a statistical framework that determines the number and area of fields of view necessary to accurately identify all cell phenotypes that are part of a tissue. This strategy was then used on imaging mass cytometry data to identify a measurement of tissue spatial segregation that enables optimal experimental design. See Bost et al., Nat Methods

Read EU GrantsAccess' Science Stories feature on the Polymenidou group's research.

Bacterial groups can coordinate cooperative actions through a communication process called quorum sensing (QS). However, whether individual bacteria coordinate their QS actions at the single-cell level, or whether group QS phenotypes are the sum of their noisy members, is unknown. Here, the Kümmerli group tracked the temporal commitments of individual Pseudomonas aeruginosa bacteria to the intertwined Las and Rhl-QS systems, from low to high population density. QS gene expression was shown to be noisy, with heterogeneity peaking during the build-up phase of QS. Discrete subgroups of cells formed that transiently segregated into two gene expression states: low Las-receptor expressers and high Las-receptor expressers. Over time, gene expression activities converged with all cells fully committing to QS. Using general mathematical models, the Kümmerli group was able to show that molecular resource limitations during the initiation phase of regulatory cascades can mechanistically trigger gene expression segregation across cells. This mechanism can operate as a built-in brake enabling a temporary bet-hedging strategy in unpredictable environments. Together, this work reveals that studying the behavior of bacterial individuals is key to understanding emergent collective actions at the group level.

See Jayakumar et al., Current Biology

This work introduces both the concept and use cases of digital twins in medicine, frames the debate on their ethical, legal and societal implications through the lens of related health digital technologies, machine learning and personalized medicine, and maps ethical challenges stemming from those. Finally, the authors lay out how digital twins may change and challenge the future practice of medicine.

See Iqbal, Krauthammer and Biller-Andorno, The Journal of Law, Medicine & Ethics

 

Current treatment planning of patients diagnosed with a brain tumor could benefit from assessment of the spatial distribution of tumor cell concentration. Magnetic resonance imaging (MRI) can reveal areas of high cell density in gliomas, but areas of low cell concentration, which can serve as a source for the secondary appearance of the tumor after treatment, are not detected. To estimate tumor cell densities beyond the visible boundaries of the lesion, numerical simulations of tumor growth could complement imaging information by providing estimates of full spatial distributions of tumor cells. In this work, the Menze group introduces Learn-Morph-Infer: a deep learning based methodology for inferring the patient-specific spatial distribution of brain tumors from T1Gd and FLAIR MRI medical scans. The method achieves real-time performance in the order of minutes on widely available hardware and the compute time is stable across tumor models of different complexity. As such, the proposed inverse solution approach allows for clinical translation of brain tumor personalization and could also be adopted to other scientific and engineering domains.

See Ezhov et al., Med Image Anal.
 

Pseudomonas aeruginosa uses quorum sensing (QS) to coordinate expression of traits required for growth and virulence in the context of infections. Despite its importance for bacterial fitness, the QS regulon appears to be a common mutational target during long-term adaptation of P. aeruginosa in the host, natural environments, and experimental evolutions. By examining mutation types in the three QS regulons of 61 experimentally evolved QS mutants, the Kümmerli group found that mutations involving the master regulator, LasR, resulted in an almost complete breakdown of QS, whereas mutations in RhlR and PqsR resulted in changes in regulon structure and the QS-regulated trait profile. Beyond affecting the plasticity and diversity of evolved populations, these mutations might also impact bacterial fitness and virulence during infections.

See Jayakumar, Figueiredo & Kümmerli, mSystems

The quantification of hemodynamics using 4D-flow magnetic resonance imaging (MRI) data requires an adequate spatio-temporal vector field resolution at a low noise level. Here, the Menze group provides a deep convolutional neural network (CNN) that learns the inter-scale relationship of the velocity vector map and leverages an efficient residual learning scheme to make it computationally feasible. A detailed comparative study between the proposed super-resolution and the conventional cubic B-spline based vector-field super-resolution shows that the new method not only improves the peak-velocity to noise ratio of the flow field by 10% and 30% for in vivo cardiovascular and cerebrovascular data, respectively, for 4 × super-resolution, but also offers 10x faster inference over the state-of-the-art.

See Shit et al., Front. Artif. Intell
 

 

 

An open question in microbiology is whether seemingly coordinated group-level responses, such as biofilm formation, actually mirror what individual cells do. To tackle this question, the Kümmerli group used single-cell microscopy to simultaneously quantify the investment of individual P. aeruginosa cells into two public goods. Using gene expression as a proxy for investment, these bacterial cells initially show no coordination in siderophore investment, but rather high heterogeneity and bi-modality. However, with increasing cell density, gene expression becomes more homogenised across cells, with positive associations in siderophore gene expression across cells and with cell-to-cell variation correlating with cellular metabolic states. This suggests that siderophore-mediated signalling aligns behavior of individual bacteria over time and spurs a coordinated three-phase siderophore investment cycle, covering the time spans from low to high population density, steered by the various interconnected regulatory mechanisms governing siderophore synthesis.

See Mridha & Kümmerli, Commun. Biol.

Read EU GrantsAccess' Science Stories feature on the Bodenmiller group's research.

Regulatory T cells (Tregs) curb excessive immune responses and dampen inflammation. Furthermore, in non-lymphoid tissues, Tregs promote tissue homeostasis, regeneration and repair. Profound understanding of the tissue-specific adaptations and functions of these Tregs might pave the way for therapeutic approaches targeting their regenerative role. In this review, the Joller group outlines their current understanding of how Tregs migrate into peripheral tissues and the factors required for their maintenance at these sites. Furthermore, tissue-specific adaptations of Tregs at barrier and immuno-privileged sites are discussed, as well as the mechanisms that regulate Tregs' function within these organs. Finally, this work outlines what is known about the interactions of Tregs with non-immune cells in the different peripheral tissues at steady state and upon challenge or tissue damage.

See Estrada Brull, Panetti & Joller, Frontiers in Immunology

In routine clinical quantitative magnetic resonance imaging (MRI), motion artifacts affect parameter estimation and thus data quality. In this work, the Menze group presents a multiscale 3D convolutional neural network (CNN) that learns the nonlinear relationship between motion-influenced quantitative parameter maps and the residual error to their motion-free reference. A physically informed simulation is proposed for supervised model training, which generates independent paired data sets from a priori motion-free data. The proposed motion correction CNN outperforms the current state-of-the-art and reliably provides high, clinically relevant image quality for mild to pronounced patient motion.

See Pirkl et al., Med. Image Anal. 
 

An open question in microbiology is whether natural selection can favour division of labour where subpopulations or species specialise in the production of a single public good, whilst sharing the complementary goods at the group level. In this work, the Kümmerli group explored the conditions under which specialisation can lead to division of labour. By growing engineered specialists of the bacterium P. aeruginosa, each of which which only produce one of two siderophores, at different mixing ratios under varying levels of iron limitation, they could show that enforcing specialisation with regard to siderophore production does not lead to beneficial division of labour in P. aeruginosa. Rather, this enforced specialisation leads to the stable co-existence of the two specialists through mutual cheating. Mridha and Kümmerli propose that in generalists, natural selection might favour fine-tuned regulatory mechanisms over division of labour, because in fluctuating environments, this fine-tuning allows generalists to maintain the flexibility to adequately adjust public good investments.

See Mridha & Kümmerli, J Evol Biol.

To predict the effects of immunotherapy in cancer, better spatial understanding of the tumor microenvironment (TME) is needed. In this work, the Bodenmiller group characterised chemokine expression and function in samples from 69 patients with metastatic melanoma. Using multiplexed mass cytometry-based imaging of protein markers and RNA transcripts, they found that CXCL9 and CXCL10 were present in patches with CXCL13+ exhausted T cells, suggesting that they recruited B cells and aided in the formation of tertiary lymphoid structures (TLS) in melanoma. TLS had a spatial enrichment of naïve and naïve-like T cells, which are involved in anti-tumor responses. Together, this study highlights the strength of targeted RNA and protein co-detection to analyse TME based on chemokine expression and suggests that the formation of tertiary lymphoid structures may be accompanied by naïve and naïve-like T cell recruitment, which may contribute to anti-tumor activity.

See Hoch, Schulz et al., Sci Immunol

Please have a look at the UZH News and  ERC Announcement

 

In various spinal disorders, a biomechanical load analysis of the spine in the upright position is useful to understand the underlying causes of the disorder and to help guide therapy. Despite the complex 3D shape of the human spine, this analysis is typically performed using 2D radiographs. Here, the Menze group proposes a novel deep neural network architecture, which takes orthogonal 2D radiographs and infers the spine’s 3D posture using vertebral shape priors to reconstruct the 3D spinal pose in an upright standing position.

See Bayat et al., Tomography
 

See Sahadevan, Pérez-Berlanga & Polymenidou,Methods Mol. Biol.

Please find the full announcement here: Stiftung Pfizer Forschungspreis 
and the UZH News article here.