Benchmarking feature selection methods for compressing image information in high-content screening.

Biopharmaceutical drug discovery, as of today is a highly automated, high throughput endeavor, where many screening technologies produce a high-dimensional measurement per sample. A striking example is High Content Screening (HCS), which utilizes automated microscopy to systematically access the wealth of information contained in biological assays. Exploiting HCS to its full potential traditionally requires extracting a high number of features from the images to capture as much information as possible, then performing algorithmic analysis and complex data visualization in order to render this high-dimensional data into an interpretable and instructive information for guiding drug development. In this process, automated feature selection methods condense the feature set to reduce non-useful or redundant information and render it more meaningful. We compare 12 state-of-the-art feature selection methods (both supervised and unsupervised) by systematically testing them on two HCS datasets from drug screening imaging assays of high practical relevance. Considering as evaluation metrics standard plate-, assay- or compound statistics on the final results, we assess the generalizability and importance of the selected features by use of automated machine learning (AutoML) to achieve an unbiased evaluation across methods. Results provide practical guidance on experiment design, optimal sizing of a reduced feature set and choice of feature selection method, both in situations where useful experimental control states are available (enabling use of supervised algorithms) or where such controls are unavailable, using unsupervised techniques.

Deep Learning-Based HCS Image Analysis for the Enterprise.

Drug discovery programs are moving increasingly toward phenotypic imaging assays to model disease-relevant pathways and phenotypes in vitro. These assays offer richer information than target-optimized assays by investigating multiple cellular pathways simultaneously and producing multiplexed readouts. However, extracting the desired information from complex image data poses significant challenges, preventing broad adoption of more sophisticated phenotypic assays. Deep learning-based image analysis can address these challenges by reducing the effort required to analyze large volumes of complex image data at a quality and speed adequate for routine phenotypic screening in pharmaceutical research. However, while general purpose deep learning frameworks are readily available, they are not readily applicable to images from automated microscopy. During the past 3 years, we have optimized deep learning networks for this type of data and validated the approach across diverse assays with several industry partners. From this work, we have extracted five essential design principles that we believe should guide deep learning-based analysis of high-content images and multiparameter data: (1) insightful data representation, (2) automation of training, (3) multilevel quality control, (4) knowledge embedding and transfer to new assays, and (5) enterprise integration. We report a new deep learning-based software that embodies these principles, Genedata Imagence, which allows screening scientists to reliably detect stable endpoints for primary drug response, assess toxicity and safety-relevant effects, and discover new phenotypes and compound classes. Furthermore, we show how the software retains expert knowledge from its training on a particular assay and successfully reapplies it to different, novel assays in an automated fashion.

Know When You Don't Know: A Robust Deep Learning Approach in the Presence of Unknown Phenotypes.

Deep convolutional neural networks show outstanding performance in image-based phenotype classification given that all existing phenotypes are presented during the training of the network. However, in real-world high-content screening (HCS) experiments, it is often impossible to know all phenotypes in advance. Moreover, novel phenotype discovery itself can be an HCS outcome of interest. This aspect of HCS is not yet covered by classical deep learning approaches. When presenting an image with a novel phenotype to a trained network, it fails to indicate a novelty discovery but assigns the image to a wrong phenotype. To tackle this problem and address the need for novelty detection, we use a recently developed Bayesian approach for deep neural networks called Monte Carlo (MC) dropout to define different uncertainty measures for each phenotype prediction. With real HCS data, we show that these uncertainty measures allow us to identify novel or unclear phenotypes. In addition, we also found that the MC dropout method results in a significant improvement of classification accuracy. The proposed procedure used in our HCS case study can be easily transferred to any existing network architecture and will be beneficial in terms of accuracy and novelty detection.

Developing Deep Learning Applications for Life Science and Pharma Industry.

Deep Learning has boosted artificial intelligence over the past 5 years and is seen now as one of the major technological innovation areas, predicted to replace lots of repetitive, but complex tasks of human labor within the next decade. It is also expected to be 'game changing' for research activities in pharma and life sciences, where large sets of similar yet complex data samples are systematically analyzed. Deep learning is currently conquering formerly expert domains especially in areas requiring perception, previously not amenable to standard machine learning. A typical example is the automated analysis of images which are typically produced en-masse in many domains, e. g., in high-content screening or digital pathology. Deep learning enables to create competitive applications in so-far defined core domains of 'human intelligence'. Applications of artificial intelligence have been enabled in recent years by (i) the massive availability of data samples, collected in pharma driven drug programs (='big data') as well as (ii) deep learning algorithmic advancements and (iii) increase in compute power. Such applications are based on software frameworks with specific strengths and weaknesses. Here, we introduce typical applications and underlying frameworks for deep learning with a set of practical criteria for developing production ready solutions in life science and pharma research. Based on our own experience in successfully developing deep learning applications we provide suggestions and a baseline for selecting the most suited frameworks for a future-proof and cost-effective development.

Discrimination between random and non-random processes in early bacterial colonization on biomaterial surfaces: application of point pattern analysis.

The dynamics of adhesion and growth of bacterial cells on biomaterial surfaces play an important role in the formation of biofilms. The surface properties of biomaterials have a major impact on cell adhesion processes, eg the random/non-cooperative adhesion of bacteria. In the present study, the spatial arrangement of Escherichia coli on different biomaterials is investigated in a time series during the first hours after exposure. The micrographs are analyzed via an image processing routine and the resulting point patterns are evaluated using second order statistics. Two main adhesion mechanisms can be identified: random adhesion and non-random processes. Comparison with an appropriate null-model quantifies the transition between the two processes with statistical significance. The fastest transition to non-random processes was found to occur after adhesion on PTFE for 2-3 h. Additionally, determination of cell and cluster parameters via image processing gives insight into surface influenced differences in bacterial micro-colony formation.

Reproducible biofilm cultivation of chemostat-grown Escherichia coli and investigation of bacterial adhesion on biomaterials using a non-constant-depth film fermenter.

Biomaterials-associated infections are primarily initiated by the adhesion of microorganisms on the biomaterial surfaces and subsequent biofilm formation. Understanding the fundamental microbial adhesion mechanisms and biofilm development is crucial for developing strategies to prevent such infections. Suitable in vitro systems for biofilm cultivation and bacterial adhesion at controllable, constant and reproducible conditions are indispensable. This study aimed (i) to modify the previously described constant-depth film fermenter for the reproducible cultivation of biofilms at non-depth-restricted, constant and low shear conditions and (ii) to use this system to elucidate bacterial adhesion kinetics on different biomaterials, focusing on biomaterials surface nanoroughness and hydrophobicity. Chemostat-grown Escherichia coli were used for biofilm cultivation on titanium oxide and investigating bacterial adhesion over time on titanium oxide, poly(styrene), poly(tetrafluoroethylene) and glass. Using chemostat-grown microbial cells (single-species continuous culture) minimized variations between the biofilms cultivated during different experimental runs. Bacterial adhesion on biomaterials comprised an initial lag-phase I followed by a fast adhesion phase II and a phase of saturation III. With increasing biomaterials surface nanoroughness and increasing hydrophobicity, adhesion rates increased during phases I and II. The influence of materials surface hydrophobicity seemed to exceed that of nanoroughness during the lag-phase I, whereas it was vice versa during adhesion phase II. This study introduces the non-constant-depth film fermenter in combination with a chemostat culture to allow for a controlled approach to reproducibly cultivate biofilms and to investigate bacterial adhesion kinetics at constant and low shear conditions. The findings will support developing and adequate testing of biomaterials surface modifications eventually preventing biomaterial-associated infections.

Quantification of the interaction between biomaterial surfaces and bacteria by 3-D modeling.

It is general knowledge that bacteria/surface interactions depend on the surface topography. However, this well-known dependence has so far not been included in the modeling efforts. We propose a model for calculating interaction energies between spherical bacteria and arbitrarily structured 3-D surfaces, combining the Derjaguin, Landau, Verwey, Overbeek theory and an extended surface element integration method. The influence of roughness on the interaction (for otherwise constant parameters, e.g. surface chemistry, bacterial hydrophobicity) is quantified, demonstrating that common experimental approaches which consider amplitude parameters of the surface topography but which ignore spacing parameters fail to adequately describe the influence of surface roughness on bacterial adhesion. The statistical roughness profile parameters arithmetic average height (representing an amplitude parameter) and peak density (representing a spacing parameter) both exert a distinct influence on the interaction energy. The influence of peak density on the interaction energy increases with decreasing arithmetic average height and contributes significantly to the total interaction energy with an arithmetic average height below 70 nm. With the aid of the proposed model, different sensitivity ranges of the interaction between rough surfaces and bacteria are identified. On the nanoscale, the spacing parameter of the surface dominates the interaction, whereas on the microscale the amplitude parameter adopts the governing role.

Evaluation of wettability and surface energy of native Nitinol surfaces in relation to hemocompatibility.

In the present study the dependence of Nitinol contact angles and surface energy on surface treatment is explored in order to better understand the material hemocompatibility that was evaluated in our previous studies. It is found that in the group of surfaces: (1) mechanically polished, (2) additionally heat treated, (3) chemically etched, and (4) additionally boiled in water, and (5) further heat treated, the contact angle could vary in the 50°-80° hydrophobic range and the total surface free energy in the 34-53 mN/m range. The polar surface energy, varying from 5 to 29 mN/m, constitutes a decisive contribution to the total energy change, and it seems to be a direct function of the Nitinol surface chemistry. Based on the complex analysis of surface energy together with the earlier results on electrochemistry and hemocompatibility it is concluded that the alteration of the polar component of surface energy and thrombogenicity is due to changes of the electron-acceptor/electron-donor character of native Nitinol surfaces during surface treatments.

Fibrinogen adsorption on biomaterials--a numerical study.

A model for the adsorption of fibrinogen or, in general, non-globular shaped proteins on solid surfaces are presented. Two-dimensional cellular automata simulations of the adsorption of fibrinogen on two different surfaces were performed. The model includes mass transfer toward the surface, adsorption of fibrinogen molecules, and surface diffusion mechanisms for both fibrinogen molecules and clusters. We show that the major physical processes are represented in the recent model. Particularly, the influence of the surface hydrophobicity on the behavior of fibrinogen. Atomic force microscopy images of fibrinogen adsorption on Si model surfaces with different hydrophobicity are compared to the results.