Search

Metrics optimized in complex machine learning tasks are often selected in an ad-hoc manner. It is unknown how they align with human expert perception. We explore the correlations between established quantitative segmentation quality metrics and qualitative evaluations by professionally trained human raters. Therefore, we conduct psychophysical experiments for two complex biomedical semantic segmentation problems. We discover that current standard metrics and loss functions correlate only moderately with the segmentation quality assessment of experts. Importantly, this effect is particularly pronounced for clinically relevant structures, such as the enhancing tumor compartment of glioma in brain magnetic resonance and grey matter in ultrasound imaging. It is often unclear how to optimize abstract metrics, such as human expert perception, in convolutional neural network (CNN) training. To cope with this challenge, we propose a novel strategy employing techniques of classical statistics to create complementary compound loss functions to better approximate human expert perception. Across all rating experiments, human experts consistently scored computer-generated segmentations better than the human-curated reference labels. Our results, therefore, strongly question many current practices in medical image segmentation and provide meaningful cues for future research., Accepted for publication at the Journal of Machine Learning for Biomedical Imaging (MELBA) https://melba-journal.org/2023:002

Large language models (LLMs) have greatly enhanced our ability to understand biology and chemistry. Yet, relatively few robust methods have been reported for structure-based drug discovery. Highly precise biomolecule-ligand interaction datasets are urgently needed in particular for LLMs, that require extensive training data. We present MISATO, the first dataset that combines quantum mechanics properties of small molecules and associated molecular dynamics simulations of about 20000 experimental protein-ligand complexes. Starting from the PDBbind dataset, semi-empirical quantum mechanics was used to systematically refine these structures. The largest collection to date of molecular dynamics traces of protein-ligand complexes in explicit water are included, accumulating to 170 μs. We give ML baseline models and simple Python data loaders, and aim to foster a thriving community around MISATO (https://github.com/t7morgen/misato-dataset). An easy entry point for ML experts is provided without the need of deep domain expertise to enable the next generation of drug discovery AI models.

Echocardiography, a rapid and cost-effective imaging technique, assesses cardiac function and structure. Despite its popularity in cardiovascular medicine and clinical research, image-derived phenotypic measurements are manually performed, requiring expert knowledge and training. Notwithstanding great progress in deep-learning applications in small animal echocardiography, the focus has so far only been on images of anesthetized rodents. We present here a new algorithm specifically designed for echocardiograms acquired in conscious mice called Echo2Pheno, an automatic statistical learning workflow for analyzing and interpreting high-throughput non-anesthetized transthoracic murine echocardiographic images in the presence of genetic knockouts. Echo2Pheno comprises a neural network module for echocardiographic image analysis and phenotypic measurements, including a statistical hypothesis-testing framework for assessing phenotypic differences between populations. Using 2159 images of 16 different knockout mouse strains of the German Mouse Clinic, Echo2Pheno accurately confirms known cardiovascular genotype–phenotype relationships (e.g., Dystrophin) and discovers novel genes (e.g., CCR4-NOT transcription complex subunit 6-like, Cnot6l, and synaptotagmin-like protein 4, Sytl4), which cause altered cardiovascular phenotypes, as verified by H&E-stained histological images. Echo2Pheno provides an important step toward automatic end-to-end learning for linking echocardiographic readouts to cardiovascular phenotypes of interest in conscious mice. Graphical abstract

BackgroundMany automatic approaches to brain tumor segmentation employ multiple magnetic resonance imaging (MRI) sequences. The goal of this project was to compare different combinations of input sequences to determine which MRI sequences are needed for effective automated brain metastasis (BM) segmentation.MethodsWe analyzed preoperative imaging (T1-weighted sequence ± contrast-enhancement (T1/T1-CE), T2-weighted sequence (T2), and T2 fluid-attenuated inversion recovery (T2-FLAIR) sequence) from 333 patients with BMs from six centers. A baseline 3D U-Net with all four sequences and six U-Nets with plausible sequence combinations (T1-CE, T1, T2-FLAIR, T1-CE+T2-FLAIR, T1-CE+T1+T2-FLAIR, T1-CE+T1) were trained on 239 patients from two centers and subsequently tested on an external cohort of 94 patients from four centers.ResultsThe model based on T1-CE alone achieved the best segmentation performance for BM segmentation with a median Dice similarity coefficient (DSC) of 0.96. Models trained without T1-CE performed worse (T1-only: DSC = 0.70 and T2-FLAIR-only: DSC = 0.72). For edema segmentation, models that included both T1-CE and T2-FLAIR performed best (DSC = 0.93), while the remaining four models without simultaneous inclusion of these both sequences reached a median DSC of 0.81-0.89.ConclusionsA T1-CE-only protocol suffices for the segmentation of BMs. The combination of T1-CE and T2-FLAIR is important for edema segmentation. Missing either T1-CE or T2-FLAIR decreases performance. These findings may improve imaging routines by omitting unnecessary sequences, thus allowing for faster procedures in daily clinical practice while enabling optimal neural network-based target definitions.

Metabolic rewiring is essential to enable cancer onset and progression. One important metabolic pathway that is often hijacked by cancer cells is the one-carbon cycle, in which the third carbon of serine is oxidized to formate. We have previously shown that formate production in cancer cells often exceeds the anabolic demand, resulting in formate overflow. Furthermore, we observed that high extracellular formate promotes thein vitroinvasiveness of glioblastoma (GBM) cells. However, additional data supporting thisin vitroobservation and mechanistic details remained elusive so far.In the present study, we now demonstrate that inhibition of formate overflow results in a decreased invasiveness of GBM cellsex vivoandin vivo. Additionally, we observed that exposure to exogeneous formate can induce a transiently stable pro-invasive phenotype that results in increased metastasis formationin vivo. All in all, these results suggest that a local formate increase within the tumor microenvironment may be one factor that can promote cancer cell motility and dissemination.Mechanistically, we uncover a previously undescribed interplay where formate acts as a trigger to alter fatty acid metabolism and matrix metalloproteinase (MMP) activity which in turn impacts cancer cell invasiveness. We thus highlight the role of formate as a pro-invasive metabolite. Gaining a deeper understanding of formate overflow and how it promotes invasion in cancer, may open new therapeutic opportunities to prevent cancer cell dissmination.

Interstitial lung disease (ILD) is a heterogenous group of lung disorders where destruction and incomplete regeneration of the lung parenchyma often results in persistent architectural distortion of the pulmonary scaffold. Continuous mesenchyme-centered, disease-relevant signaling likely initiates and perpetuates the fibrotic remodeling process, specifically targeting the epithelial cell compartment, thereby destroying the gas exchange area.With the aim of identifying functionally relevant mediators of the lung mesenchymal-epithelial crosstalk that hold potential as new targets for therapeutic strategies, we developed a 3D organoid co-culture model based on human induced pluripotent stem cell-derived alveolar epithelial type 2 cells that form alveolar organoids in presence of lung fibroblasts from ILD patients as well as a control cell line (IMR-90). While organoid formation capacity and size was comparable in the presence of ILD or control lung fibroblasts, metabolic activity was significantly increased in ILD co-cultures. Alveolar organoids cultured with ILD fibroblasts further demonstrated reduced stem cell function as reflected by reducedSurfactant Protein Cgene expression together with an aberrant basaloid-prone differentiation program indicated by elevatedCadherin 2, Bone Morphogenic Protein 4andVimentintranscription.In order to screen for key mediators of the misguided mesenchymal-to-epithelial crosstalk with a focus on disease-relevant inflammatory processes, we used mass spectrometry and characterized the secretome of end stage ILD lung fibroblasts in comparison to non-chronic lung disease (CLD) patient fibroblasts. Out of the over 2000 proteins detected by this experimental approach, 47 proteins were differentially abundant comparing ILD and non-CLD fibroblast secretome.The ILD secretome profile was dominated by chemokines, includingCXCL1, CXCL3, andCXCL8, interfering with growth factor signaling orchestrated byInterleukin 11 (IL11), steering fibrogenic cell-cell communication, and proteins regulating extracellular matrix remodeling including epithelial-to-mesenchymal transition. When in turn treating alveolar organoids withIL11, we recapitulated the co-culture results obtained with primary ILD fibroblasts including changes in metabolic activity.In summary, we identified mediators likely contributing to the disease-perpetuating mesenchymal-to-epithelial crosstalk in ILD. In our alveolar organoid co-cultures, we were able to highlight the importance of fibroblast-initiated aberrant epithelial differentiation and confirmedIL11as a key player in ILD pathogenesis by unbiased ILD fibroblast secretome analysis.Abstract Figure

Targeted spatial transcriptomics methods capture the topology of cell types and states in tissues at single cell- and subcellular resolution by measuring the expression of a predefined set of genes. The selection of an optimal set of probed genes is crucial for capturing and interpreting the spatial signals present in a tissue. However, current selections often rely on marker genes, precluding them from detecting continuous spatial signals or novel states. We present Spapros, an end-to-end probe set selection pipeline that optimizes both probe set specificity for cell type identification and within-cell-type expression variation to resolve spatially distinct populations while taking into account prior knowledge, as well as probe design and expression constraints. To facilitate data analysis and interpretation, Spapros also provides rules for cell type identification. We evaluated Spapros by selecting probes on 6 different data sets and built an evaluation pipeline with 12 quality metrics to find that Spapros outperforms other selection approaches in both cell type recovery and recovering expression variation beyond cell types. Furthermore, we used Spapros to design a SCRINSHOT experiment of adult lung tissue to demonstrate how probes selected with Spapros identify cell types of interest and detect spatial variation even within cell types. Spapros enables optimal probe set selection, probe set evaluation, and probe design, as a freely available Python package.

Pulmonary delivery is not only of increasing importance to exploit local therapy for lung diseases like COVID-19, but also plays an essential role in interpreting health impact of inhaled hazardous (nano-)materials. Deciphering of biodistribution, biokinetics, and cell-mediated immunity of pulmonary delivered nanoparticles (NPs) remains insufficient. We applied state-of-the-art multimodal imaging and deep learning pipelines to elucidate the intricate deposition profiles and the fate of biopersistent NPs in the lung within cellular resolution. Spatially resolved delivery characteristics like inter- and intra-acinar distribution pattern and regional NP dosage were revealed by tissue-clearing light-sheet fluorescence microscopy for six common routes of pulmonary delivery, i.e., intranasal & oropharyngeal aspiration, intubated instillation, microsprayer application, ventilator-assisted & nose-only aerosol inhalation. Artificial intelligence-driven entire airway 3D segmentation (convolutional neural network) enabled accurate dosimetry of bronchial versus acinar/alveolar NP dose (B/A: 0.04 to 0.18 depending on delivery route) rather than the conventionally simplified 2D central-to-peripheral ratio (C/P: 1.32 to 3.79). Moreover, longitudinal 3D imaging revealed similar intra-acinar NP kinetics independent of delivery route. Inhaled NP relocation from the alveolar-epithelium to interstitium was traditionally claimed as the consequence of passive diffusion/transportation of NPs. This study, however, revealed that acinar NP transport can be attributed to the phagocytosis and migration of tissue-resident macrophages (TRMs) as evidenced by complementary analytical approaches (intravital microscopy, ex vivo 4D live-tissue microscopy, in situ immunofluorescence, ex vivo NP-lung interaction model, flow cytometry). Overall, this study elucidates holistic NP-lung delivery and biokinetics features and TRM immunity on the fate of biopersistent NPs.

CRISPR interference (CRISPRi), the targeting of a catalytically dead Cas protein to block transcription, is the leading technique to silence gene expression in bacteria. However, design rules for CRISPRi remain poorly defined, limiting predictable design for gene interrogation, pathway manipulation, and high-throughput screens. Here we develop a best-in-class prediction algorithm for guide silencing efficiency by systematically investigating factors influencing guide depletion in multiple genome-wide essentiality screens, with the surprising discovery that gene-specific features such as transcriptional activity substantially impact prediction of guide activity. Accounting for these features as part of algorithm development allowed us to develop a mixed-effect random forest regression model that provides better estimates of guide efficiency than existing methods, as demonstrated in an independent saturating screen. We further applied methods from explainable AI to extract interpretable design rules from the model, such as sequence preferences in the vicinity of the PAM distinct from those previously described for genome engineering applications. Our approach provides a blueprint for the development of predictive models for CRISPR technologies where only indirect measurements of guide activity are available.

We present DeepVesselNet, an architecture tailored to the challenges faced when extracting vessel networks or trees and corresponding features in 3-D angiographic volumes using deep learning. We discuss the problems of low execution speed and high memory requirements associated with full 3-D convolutional networks, high-class imbalance arising from the low percentage of vessel voxels, and unavailability of accurately annotated training data - and offer solutions as the building blocks of DeepVesselNet. First, we formulate 2-D orthogonal cross-hair filters which make use of 3-D context information at a reduced computational burden. Second, we introduce a class balancing cross-entropy loss function with false positive rate correction to handle the high-class imbalance and high false positive rate problems associated with existing loss functions. Finally, we generate synthetic dataset using a computational angiogenesis model capable of generating vascular trees under physiological constraints on local network structure and topology and use these data for transfer learning. DeepVesselNet is optimized for segmenting and analyzing vessels, and we test the performance on a range of angiographic volumes including clinical MRA data of the human brain, as well as X-ray tomographic microscopy scans of the rat brain. Our experiments show that, by replacing 3-D filters with cross-hair filters in our network, we achieve over 23% improvement in speed, lower memory footprint, lower network complexity which prevents overfitting and comparable accuracy (with a Cox-Wilcoxon paired sample significance test p-value of 0.07 when compared to full 3-D filters). Our class balancing metric is crucial for training the network and transfer learning with synthetic data is an efficient, robust, and very generalizable approach leading to a network that excels in a variety of angiography segmentation tasks., 13 pages