deepFPlearn +: enhancing toxicity prediction across the chemical universe using graph neural networks.

SUMMARY: Sophisticated approaches for the in silico prediction of toxicity are required to support the risk assessment of chemicals. The number of chemicals on the global chemical market and the speed of chemical innovation stand in massive contrast to the capacity for regularizing chemical use. We recently proved our ready-to-use application deepFPlearn as a suitable approach for this task. Here, we present its extension deepFPlearn+ incorporating (i) a graph neural network to feed our AI with a more sophisticated molecular structure representation and (ii) alternative train-test splitting strategies that involve scaffold structures and the molecular weights of chemicals. We show that the GNNs outperform the previous model substantially and that our models can generalize on unseen data even with a more robust and challenging test set. Therefore, we highly recommend the application of deepFPlearn+ on the chemical inventory to prioritize chemicals for experimental testing or any chemical subset of interest in monitoring studies. AVAILABILITY AND IMPLEMENTATION: The software is compatible with python 3.6 or higher, and the source code can be found on our GitHub repository: https://github.com/yigbt/deepFPlearn. The data underlying this article are available in Zenodo, and can be accessed with the link below: https://zenodo.org/record/8146252. Detailed installation guides via Docker, Singularity, and Conda are provided within the repository for operability across all operating systems.

Error quantification in multi-parameter mapping facilitates robust estimation and enhanced group level sensitivity.

Multi-Parameter Mapping (MPM) is a comprehensive quantitative neuroimaging protocol that enables estimation of four physical parameters (longitudinal and effective transverse relaxation rates R1 and R2*, proton density PD, and magnetization transfer saturation MTsat) that are sensitive to microstructural tissue properties such as iron and myelin content. Their capability to reveal microstructural brain differences, however, is tightly bound to controlling random noise and artefacts (e.g. caused by head motion) in the signal. Here, we introduced a method to estimate the local error of PD, R1, and MTsat maps that captures both noise and artefacts on a routine basis without requiring additional data. To investigate the method's sensitivity to random noise, we calculated the model-based signal-to-noise ratio (mSNR) and showed in measurements and simulations that it correlated linearly with an experimental raw-image-based SNR map. We found that the mSNR varied with MPM protocols, magnetic field strength (3T vs. 7T) and MPM parameters: it halved from PD to R1 and decreased from PD to MTsat by a factor of 3-4. Exploring the artefact-sensitivity of the error maps, we generated robust MPM parameters using two successive acquisitions of each contrast and the acquisition-specific errors to down-weight erroneous regions. The resulting robust MPM parameters showed reduced variability at the group level as compared to their single-repeat or averaged counterparts. The error and mSNR maps may better inform power-calculations by accounting for local data quality variations across measurements. Code to compute the mSNR maps and robustly combined MPM maps is available in the open-source hMRI toolbox.

AI for predicting chemical-effect associations at the chemical universe level-deepFPlearn.

Many chemicals are present in our environment, and all living species are exposed to them. However, numerous chemicals pose risks, such as developing severe diseases, if they occur at the wrong time in the wrong place. For the majority of the chemicals, these risks are not known. Chemical risk assessment and subsequent regulation of use require efficient and systematic strategies. Lab-based methods-even if high throughput-are too slow to keep up with the pace of chemical innovation. Existing computational approaches are designed for specific chemical classes or sub-problems but not usable on a large scale. Further, the application range of these approaches is limited by the low amount of available labeled training data. We present the ready-to-use and stand-alone program deepFPlearn that predicts the association between chemical structures and effects on the gene/pathway level using a combined deep learning approach. deepFPlearn uses a deep autoencoder for feature reduction before training a deep feed-forward neural network to predict the target association. We received good prediction qualities and showed that our feature compression preserves relevant chemical structural information. Using a vast chemical inventory (unlabeled data) as input for the autoencoder did not reduce our prediction quality but allowed capturing a much more comprehensive range of chemical structures. We predict meaningful-experimentally verified-associations of chemicals and effects on unseen data. deepFPlearn classifies hundreds of thousands of chemicals in seconds. We provide deepFPlearn as an open-source and flexible tool that can be easily retrained and customized to different application settings at https://github.com/yigbt/deepFPlearn.

Transforming Growth Factor Beta 3-Loaded Decellularized Equine Tendon Matrix for Orthopedic Tissue Engineering.

Transforming growth factor beta 3 (TGFß3) promotes tenogenic differentiation and may enhance tendon regeneration in vivo. This study aimed to apply TGFß3 absorbed in decellularized equine superficial digital flexor tendon scaffolds, and to investigate the bioactivity of scaffold-associated TGFß3 in an in vitro model. TGFß3 could effectively be loaded onto tendon scaffolds so that at least 88% of the applied TGFß3 were not detected in the rinsing fluid of the TGFß3-loaded scaffolds. Equine adipose tissue-derived multipotent mesenchymal stromal cells (MSC) were then seeded on scaffolds loaded with 300 ng TGFß3 to assess its bioactivity. Both scaffold-associated TGFß3 and TGFß3 dissolved in the cell culture medium, the latter serving as control group, promoted elongation of cell shapes and scaffold contraction (p < 0.05). Furthermore, scaffold-associated and dissolved TGFß3 affected MSC musculoskeletal gene expression in a similar manner, with an upregulation of tenascin c and downregulation of other matrix molecules, most markedly decorin (p < 0.05). These results demonstrate that the bioactivity of scaffold-associated TGFß3 is preserved, thus TGFß3 application via absorption in decellularized tendon scaffolds is a feasible approach.

Growth Factor-Mediated Tenogenic Induction of Multipotent Mesenchymal Stromal Cells Is Altered by the Microenvironment of Tendon Matrix.

Age-related degenerative changes in tendon tissue represent a common cause for acute tendon pathologies. Although the regenerative potential of multipotent mesenchymal stromal cells (MSC) was reported to restore functionality in injured tendon tissue, cellular mechanisms of action remain partly unclear. Potential tenogenic differentiation of applied MSC is affected by various intrinsic and extrinsic factors. The current study presents an in vitro model to evaluate the combined extrinsic effects of decellularized equine tendon matrix, transforming growth factor beta 3 (TGFß3) and bone morphogenetic protein 12 (BMP12) on the tenogenic fate of equine adipose tissue-derived MSC. Monolayer MSC cultures supplemented with TGFß3 and BMP12 as well as MSC cultured on tendon matrix scaffolds preloaded with the growth factors were incubated for 3 and 5 days. Histological evaluation and real time reverse transcription polymerase chain reaction (RT-PCR) revealed that growth factor-mediated tenogenic induction of MSC was modified by the conditions of the surrounding microenvironment. While the gene expression pattern in monolayer cultures supplemented with TGFß3 or TGFß3 and BMP12 revealed an upregulation for collagen 1A2, collagen 3A1, tenascin c, scleraxis and mohawk ( p < 0.05 ), the presence of tendon matrix led to an upregulation of decorin and osteopontin as well as to a downregulation of smad8 ( p < 0.05). Preloading of scaffolds with either TGFß3, or with TGFß3 and BMP12 promoted a tenocyte-like phenotype and improved cell alignment. Furthermore, gene expression in scaffold culture was modulated by TGFß3 and/or BMP12, with downregulation of collagen 1A2, collagen 3A1, decorin, scleraxis, smad8 and osteopontin, whereas gene expression of tenascin c was increased. This study shows that growth factor-induced tenogenic differentiation of equine MSC is markedly altered by topographical constraints of decellularized tendon tissue in vitro. While TGFß3 represents an effective mediator for tenogenic induction, the role of BMP12 in tenogenesis may be of modulatory character and needs further evaluation.

Tenogenic Properties of Mesenchymal Progenitor Cells Are Compromised in an Inflammatory Environment.

Transplantation of multipotent mesenchymal progenitor cells is a valuable option for treating tendon disease. Tenogenic differentiation leading to cell replacement and subsequent matrix modulation may contribute to the regenerative effects of these cells, but it is unclear whether this occurs in the inflammatory environment of acute tendon disease. Equine adipose-derived stromal cells (ASC) were cultured as monolayers or on decellularized tendon scaffolds in static or dynamic conditions, the latter represented by cyclic stretching. The impact of different inflammatory conditions, as represented by supplementation with interleukin-1ß and/or tumor necrosis factor-α or by co-culture with allogeneic peripheral blood leukocytes, on ASC functional properties was investigated. High cytokine concentrations increased ASC proliferation and osteogenic differentiation, but decreased chondrogenic differentiation and ASC viability in scaffold culture, as well as tendon scaffold repopulation, and strongly influenced musculoskeletal gene expression. Effects regarding the latter differed between the monolayer and scaffold cultures. Leukocytes rather decreased ASC proliferation, but had similar effects on viability and musculoskeletal gene expression. This included decreased expression of the tenogenic transcription factor scleraxis by an inflammatory environment throughout culture conditions. The data demonstrate that ASC tenogenic properties are compromised in an inflammatory environment, with relevance to their possible mechanisms of action in acute tendon disease.

Automated detection of the choroid boundary within OCT image data using quadratic measure filters.

A novel method for the automated detection of the outer choroid boundary within spectral-domain optical coherence tomography image data, based on an image model within the space of functions of bounded variation and the application of quadratic measure filters, is presented. The same method is used for the segmentation of retinal layer boundaries and proves to be suitable even for data generated without special imaging modes and moderate line averaging. Based on the segmentations, an automated determination of the central fovea region and choroidal thickness measurements for this and two adjacent 1-mm regions are provided. The quality of the method is assessed by comparison with manual delineations performed by five trained graders. The study is based on data from 50 children of the ages 8 to 13 that were obtained in the framework of the LIFE Child study at Leipzig University.

Analysis of foveal characteristics and their asymmetries in the normal population.

The advance of optical coherence tomography (OCT) enables a detailed examination of the human retina in-vivo for clinical routine and experimental eye research. Only few investigations to date captured human foveal morphology in a large subject group on the basis of a detailed analysis employing mathematical models. However, even for important foveal characteristics unified terminology and clear definitions were not implemented so far. This might be a reason, why to this day the human fovea is considered to be a mostly symmetric and round structure. Therefore, the most important finding of this work is the detailed analysis of the asymmetric structure of the human fovea. We employed five clinically highly relevant foveal characteristics, which are derived from a previously published fovea model. For each, an accurate mathematical description is given. The presented properties include (1) mean retinal thickness inside a defined radius, (2) foveal bowl area, (3) a new, exact definition of foveal radius, (4) maximum foveal slope, and (5) the maximum height of the foveal rim. Furthermore, minimum retinal thickness was derived and analyzed. 220 strictly controlled healthy Caucasian subjects of European decent with an even distribution of age and gender were imaged with an Heidelberg Spectralis OCT. Detailed analysis demonstrated the following general results: (1) significant gender difference regarding the central foveal subfield thickness (CFST) but no significant differences for the minimum central retinal thickness, (2) a strong correlation between right and left eye of the same subject, and, as essential finding, (3) strong structural differences of the fovea form in the different anatomical directions (nasal, temporal, inferior and superior). In the analysis of the foveal asymmetry, it will be demonstrated that the foveal radius is larger in nasal and temporal direction compared to inferior and superior position. Furthermore, it will be shown that the circular fovea rather has an elliptic form with the larger axis along the nasal to temporal direction. Interestingly, the foveal slope shows a divergent behavior as the temporal direction has the smallest slope angle and both, inferior and superior angles are clearly larger than the others. The findings in this work can be used for an exact quantification of changes in early stages of various retinal diseases and as a marker for initial diagnosis.

Local Variability of Parameters for Characterization of the Corneal Subbasal Nerve Plexus.

PURPOSE: The corneal subbasal nerve plexus (SNP) offers high potential for early diagnosis of diabetic peripheral neuropathy. Changes in subbasal nerve fibers can be assessed in vivo by confocal laser scanning microscopy (CLSM) and quantified using specific parameters. While current study results agree regarding parameter tendency, there are considerable differences in terms of absolute values. The present study set out to identify factors that might account for this high parameter variability. MATERIALS AND METHODS: In three healthy subjects, we used a novel method of software-based large-scale reconstruction that provided SNP images of the central cornea, decomposed the image areas into all possible image sections corresponding to the size of a single conventional CLSM image (0.16 mm2), and calculated a set of parameters for each image section. In order to carry out a large number of virtual examinations within the reconstructed image areas, an extensive simulation procedure (10,000 runs per image) was implemented. RESULTS: The three analyzed images ranged in size from 3.75 mm2 to 4.27 mm2. The spatial configuration of the subbasal nerve fiber networks varied greatly across the cornea and thus caused heavily location-dependent results as well as wide value ranges for the parameters assessed. Distributions of SNP parameter values varied greatly between the three images and showed significant differences between all images for every parameter calculated (p < 0.001 in each case). CONCLUSIONS: The relatively small size of the conventionally evaluated SNP area is a contributory factor in high SNP parameter variability. Averaging of parameter values based on multiple CLSM frames does not necessarily result in good approximations of the respective reference values of the whole image area. This illustrates the potential for examiner bias when selecting SNP images in the central corneal area.

Parametric model for the 3D reconstruction of individual fovea shape from OCT data.

As revealed by optical coherence tomography (OCT), the shape of the fovea may vary greatly among individuals. However, none of the hitherto available mathematical descriptions comprehensively reproduces all individual characteristics such as foveal depth, slope, naso-temporal asymmetry, and others. Here, a novel mathematical approach is presented to obtain a very accurate model of the complete 3D foveal surface of an individual, by utilizing recent developments in OCT. For this purpose, a new formula was developed serving as a simple but very flexible way to represent a given fovea. An extensive description of the used model parameters, as well as, of the complete method of reconstructing a foveal surface from OCT data, is presented. Noteworthy, the formula analytically provides characteristic foveal parameters and thus allows for extensive quantification. The present approach was verified on 432 OCT scans and has proved to be able to capture the whole range of asymmetric foveal shapes with high accuracy (i.e. a mean fit error of 1.40 µm).

Nanostructured self-assembling peptides as a defined extracellular matrix for long-term functional maintenance of primary hepatocytes in a bioartificial liver modular device.

Much effort has been directed towards the optimization of the capture of in vivo hepatocytes from their microenvironment. Some methods of capture include an ex vivo cellular model in a bioreactor based liver module, a micropatterned module, a microfluidic 3D chip, coated plates, and other innovative approaches for the functional maintenance of primary hepatocytes. However, none of the above methods meet US Food and Drug Administration (FDA) guidelines, which recommend and encourage that the duration of a toxicity assay of a drug should be a minimum of 14 days, to a maximum of 90 days for a general toxicity assay. Existing innovative reports have used undefined extracellular matrices like matrigel, rigid collagen, or serum supplementations, which are often problematic, unacceptable in preclinical and clinical applications, and can even interfere with experimental outcomes. We have overcome these challenges by using integrated nanostructured self-assembling peptides and a special combination of growth factors and cytokines to establish a proof of concept to mimic the in vivo hepatocyte microenvironment pattern in vitro for predicting the in vivo drug hepatotoxicity in a scalable bioartificial liver module. Hepatocyte functionality (albumin, urea) was measured at days 10, 30, 60, and 90 and we observed stable albumin secretion and urea function throughout the culture period. In parallel, drug metabolizing enzyme biomarkers such as ethoxyresorufin-O-deethylase, the methylthiazol tetrazolium test, and the lactate dehydrogenase test were carried out at days 10, 30, 60, and 90. We noticed excellent mitochondrial status and membrane stability at 90 days of culture. Since alpha glutathione S-transferase (GST) is highly sensitive and a specific marker of hepatocyte injury, we observed significantly low alpha GST levels on all measured days (10, 30, 60, and 90). Finally, we performed the image analysis of mitochondria-cultured hepatocytes at day 90 in different biophysical parameters using confocal microscopy. We applied an automatic algorithm-based method for 3D visualization to show the classic representation of the mitochondrial distribution in double hepatocytes. An automated morphological measurement was conducted on the mitochondrial distribution in the cultured hepatocytes. Our proof of concept of a scalable bioartificial liver modular device meets FDA guidelines and may function as an alternative model of animal experimentation for pharmacological and toxicological studies involving drug metabolism, enzyme induction, transplantation, viral hepatitis, hepatocyte regeneration, and can also be used in other existing bioreactor modules for long-term culture for up to 90 days or more.

Quantitative analysis of astrogliosis in drug-dependent humans.

Drug addiction is a chronic, relapsing disease caused by neurochemical and molecular changes in the brain. In this human autopsy study qualitative and quantitative changes of glial fibrillary acidic protein (GFAP)-positive astrocytes in the hippocampus of 26 lethally intoxicated drug addicts and 35 matched controls are described. The morphological characterization of these cells reflected alterations representative for astrogliosis. But, neither quantification of GFAP-positive cells nor the Western blot analysis indicated statistical significant differences between drug fatalities versus controls. However, by semi-quantitative scoring a significant shift towards higher numbers of activated astrocytes in the drug group was detected. To assess morphological changes quantitatively, graph-based representations of astrocyte morphology were obtained from single cell images captured by confocal laser scanning microscopy. Their underlying structures were used to quantify changes in astroglial fibers in an automated fashion. This morphometric analysis yielded significant differences between the investigated groups for four different measures of fiber characteristics (Euclidean distance, graph distance, number of graph elements, fiber skeleton distance), indicating that, e.g., astrocytes in drug addicts on average exhibit significant elongation of fiber structures as well as two-fold increase in GFAP-positive fibers as compared with those in controls. In conclusion, the present data show characteristic differences in morphology of hippocampal astrocytes in drug addicts versus controls and further supports the involvement of astrocytes in human pathophysiology of drug addiction. The automated quantification of astrocyte morphologies provides a novel, testable way to assess the fiber structures in a quantitative manner as opposed to standard, qualitative descriptions.

Classifying prostate cancer malignancy by quantitative histomorphometry.

PURPOSE: Prostate cancer is routinely graded according to the Gleason grading scheme. This scheme is predominantly based on the textural appearance of aberrant glandular structures. Gleason grade is difficult to standardize and often leads to discussion due to interrater and intrarater disagreement. Thus, we investigated whether digital image based automated quantitative histomorphometry could be used to achieve a more standardized, reproducible classification outcome. MATERIALS AND METHODS: In a proof of principle study we developed a method to evaluate digitized histological images of single prostate cancer regions in hematoxylin and eosin stained sections. Preprocessed color images were subjected to color deconvolution, followed by the binarization of obtained hematoxylin related image channels. Highlighted neoplastic epithelial gland related objects were morphometrically assessed by a classifier based on 2 calculated quantitative and objective geometric measures, that is inverse solidity and inverse compactness. The procedure was then applied to the prostate cancer probes of 125 patients. Each probe was independently classified for Gleason grade 3, 4 or 5 by an experienced pathologist blinded to image analysis outcome. RESULTS: Together inverse compactness and inverse solidity were adequate discriminatory features for a powerful classifier that distinguished Gleason grade 3 from grade 4/5 histology. The classifier was robust on sensitivity analysis. CONCLUSIONS: Results suggest that quantitative and interpretable measures can be obtained from image based analysis, permitting algorithmic differentiation of prostate Gleason grades. The method must be validated in a large independent series of specimens.

Langmuir-Blodgett films of fluorinated glycolipids and polymerizable lipids and their phase separating behavior.

This paper describes the phase separating behavior of Langmuir monolayers from mixtures of different lipids that (i) either carry already a glycopeptide recognition site or can be easily modified to carry one and (ii) polymerizable lipids. To ensure demixing during compression, we used fluorinated lipids for the biological headgroups and hydrocarbon based lipids as polymerizable lipids. As a representative for a lipid monomer, which can be polymerized in the hydrophilic headgroup, a methacrylic monomer was used. As a monomer, which can be polymerized in the hydrophobic tail, a lipid with a diacetylene unit was used (pentacosadiynoic acid, PDA). The fluorinated lipids were on the one hand a perfluorinated lipid with three chains and on the other hand a partially fluorinated lipid with a T(N)-antigen headgroup. The macroscopic phase separation was observed by Brewster angle microscopy, whereas the phase separation on the nanoscale level was observed by atomic force microscopy. It turned out that all lipid mixtures showed (at least) a partial miscibility of the hydrocarbon compounds in the fluorinated compounds. This is positive for pattern formation, as it allows the formation of small demixed 2D patterned structures during crystallization from the homogeneous phase. For miscibility especially a liquid analogue phase proved to be advantageous. As lipid 3 with three fluorinated lipid chains (very stable monolayer) is miscible with the polymerizable lipids 1 and 2, it was mostly used for further investigations. For all three lipid mixtures, a phase separation on both the micrometer and the nanometer level was observed. The size of the crystalline domains could be controlled not only by varying the surface pressure but also by varying the molar composition of the mixtures. Furthermore, we showed that the binary mixture can be stabilized via UV polymerization. After polymerization and subsequent expansion of the barriers, the locked-in polymerized structures are stable even at low surface pressures (10 mN/m), where the unpolymerized mixture did not show any segregation.

Image-processing chain for a three-dimensional reconstruction of basal cell carcinomas.

Basal cell carcinoma (BCC) is the most common malignant skin cancer. For a deeper insight into the specific growth patterns of the tumorous tissue in BCC, we have focused on the development of a novel automated image-processing chain for 3D reconstruction of BCC using histopathological serial sections. For fully automatic delineation of the tumor within the tissue, we apply a fuzzy c-means segmentation method. We used a novel multi-grid form of the non-linear registration introduced by Braumann and Kuska in 2005 effectively suppressing registration runs into local minima (possibly caused by diffuse nature of the tumor). Our method was successfully applied in a proof-of-principle study for automated reconstruction.

Langmuir-Blodgett films of biocompatible poly(HPMA)-block-poly(lauryl methacrylate) and poly(HPMA)-random-poly(lauryl methacrylate): influence of polymer structure on membrane formation and stability.

Membranes based on functional biocompatible polymers can be regarded as a useful model system to study biological interactions, e.g. antibody-antigen interactions or protein polymer interactions. These model systems may give a better insight into these processes and may help to find suitable polymeric structures offering biocompatibility as well as reduced polymer protein interaction. In this respect, Langmuir-Blodgett (LB) layer formation at the air/water (A/W) interface is studied in respect to polymer architecture in this article. For this purpose, narrowly distributed N-(2-hydroxypropyl)-methacrylamide (HPMA) random and block copolymers have been prepared by the RAFT polymerization method. For random copolymers different molecular weights were prepared. As for the block copolymers also the ratio of hydrophilic and hydrophobic units was varied in order to study the influence of hydrophobic block length on collapse pressure and area. The molecular weights of all polymers were around 15 kDa and 30 kDa. In the case of block copolymers we found a direct correlation of the length of the hydrophobic block to the collapse area. Furthermore, hysteresis experiments clearly point out that block copolymers form stable LB layers. No remarkable changes in collapse pressure or area could be observed. In contrast the area occupied by random copolymers changes at each hysteresis cycle indicating a loss of polymer to the aqueous subphase. In addition the LB layers were transferred onto mica substrates. The block copolymers formed stable and defect free membranes over an area of 100 microm(2) with a roughness (rms) 1.3-1.4 A. On the contrary, membranes based on random copolymers turned out to have a higher surface roughness. Our findings clearly underline the influence of polymer structure on the LB layer formation at the A/W interface.

Diagnosis of rheumatoid arthritis using light: correction of motion artefacts and color visualization of multispectral images.

State-of-the-art image-processing methods offer new possibilities for diagnosing diseases using scattered light. The optical diagnosis of rheumatism is taken as an example to show that the diagnostic sensitivity can be improved using overlapped pseudocolored images of different wavelengths, provided that multispectral images are recorded to compensate for any motion-related artefacts that occur during examination.