High-Content Phenotypic Profiling of Drug Response Signatures across Distinct Cancer Cells

Morphological profiling for drug discovery in the era of deep learning

Morphological profiling is a valuable tool in phenotypic drug discovery. The advent of high-throughput automated imaging has enabled the capturing of a wide range of morphological features of cells or organisms in response to perturbations at the single-cell resolution. Concurrently, significant advances in machine learning and deep learning, especially in computer vision, have led to substantial improvements in analyzing large-scale high-content images at high throughput. These efforts have facilitated understanding of compound mechanism of action, drug repurposing, characterization of cell morphodynamics under perturbation, and ultimately contributing to the development of novel therapeutics. In this review, we provide a comprehensive overview of the recent advances in the field of morphological profiling. We summarize the image profiling analysis workflow, survey a broad spectrum of analysis strategies encompassing feature engineering- and deep learning-based approaches, and introduce publicly available benchmark datasets. We place a particular emphasis on the application of deep learning in this pipeline, covering cell segmentation, image representation learning, and multimodal learning. Additionally, we illuminate the application of morphological profiling in phenotypic drug discovery and highlight potential challenges and opportunities in this field.

Comparative oncology chemosensitivity assay for personalized medicine using low-coherence digital holography of dynamic light scattering from cancer biopsies

Nearly half of cancer patients who receive standard-of-care treatments fail to respond to their first-line chemotherapy, demonstrating the pressing need for improved methods to select personalized cancer therapies. Low-coherence digital holography has the potential to fill this need by performing dynamic contrast OCT on living cancer biopsies treated ex vivo with anti-cancer therapeutics. Fluctuation spectroscopy of dynamic light scattering under conditions of holographic phase stability captures ultra-low Doppler frequency shifts down to 10 mHz caused by light scattering from intracellular motions. In the comparative preclinical/clinical trials presented here, a two-species (human and canine) and two-cancer (esophageal carcinoma and B-cell lymphoma) analysis of spectral phenotypes identifies a set of drug response characteristics that span species and cancer type. Spatial heterogeneity across a centimeter-scale patient biopsy sample is assessed by measuring multiple millimeter-scale sub-samples. Improved predictive performance is achieved for chemoresistance profiling by identifying red-shifted sub-samples that may indicate impaired metabolism and removing them from the prediction analysis. These results show potential for using biodynamic imaging for personalized selection of cancer therapy.

Microsnoop: A generalist tool for microscopy image representation

Accurate profiling of microscopy images from small scale to high throughput is an essential procedure in basic and applied biological research. Here, we present Microsnoop, a novel deep learning-based representation tool trained on large-scale microscopy images using masked self-supervised learning. Microsnoop can process various complex and heterogeneous images, and we classified images into three categories: single-cell, full-field, and batch-experiment images. Our benchmark study on 10 high-quality evaluation datasets, containing over 2,230,000 images, demonstrated Microsnoop's robust and state-of-the-art microscopy image representation ability, surpassing existing generalist and even several custom algorithms. Microsnoop can be integrated with other pipelines to perform tasks such as superresolution histopathology image and multimodal analysis. Furthermore, Microsnoop can be adapted to various hardware and can be easily deployed on local or cloud computing platforms. We will regularly retrain and reevaluate the model using community-contributed data to consistently improve Microsnoop.

Revealing invisible cell phenotypes with conditional generative modeling

Biological sciences, drug discovery and medicine rely heavily on cell phenotype perturbation and microscope observation. However, most cellular phenotypic changes are subtle and thus hidden from us by natural cell variability: two cells in the same condition already look different. In this study, we show that conditional generative models can be used to transform an image of cells from any one condition to another, thus canceling cell variability. We visually and quantitatively validate that the principle of synthetic cell perturbation works on discernible cases. We then illustrate its effectiveness in displaying otherwise invisible cell phenotypes triggered by blood cells under parasite infection, or by the presence of a disease-causing pathological mutation in differentiated neurons derived from iPSCs, or by low concentration drug treatments. The proposed approach, easy to use and robust, opens the door to more accessible discovery of biological and disease biomarkers.

High throughput microscopy and single cell phenotypic image-based analysis in toxicology and drug discovery

Measuring single cell responses to the universe of chemicals (drugs, natural products, environmental toxicants etc.) is of paramount importance to human health as phenotypic variability in sensing stimuli is a hallmark of biology that is considered during high throughput screening. One of the ways to approach this problem is via high throughput, microscopy-based assays coupled with multi-dimensional single cell analysis methods. Here, we will summarize some of the efforts in this vast and growing field, focusing on phenotypic screens (e.g., Cell Painting), single cell analytics and quality control, with particular attention to environmental toxicology and drug screening. We will discuss advantages and limitations of high throughput assays with various end points and levels of complexity.

Evolution and impact of high content imaging

The field of high content imaging has steadily evolved and expanded substantially across many industry and academic research institutions since it was first described in the early 1990's. High content imaging refers to the automated acquisition and analysis of microscopic images from a variety of biological sample types. Integration of high content imaging microscopes with multiwell plate handling robotics enables high content imaging to be performed at scale and support medium- to high-throughput screening of pharmacological, genetic and diverse environmental perturbations upon complex biological systems ranging from 2D cell cultures to 3D tissue organoids to small model organisms. In this perspective article the authors provide a collective view on the following key discussion points relevant to the evolution of high content imaging: • Evolution and impact of high content imaging: An academic perspective • Evolution and impact of high content imaging: An industry perspective • Evolution of high content image analysis • Evolution of high content data analysis pipelines towards multiparametric and phenotypic profiling applications • The role of data integration and multiomics • The role and evolution of image data repositories and sharing standards • Future perspective of high content imaging hardware and software.

Application of cell painting for chemical hazard evaluation in support of screening-level chemical assessments

'Cell Painting' is an imaging-based high-throughput phenotypic profiling (HTPP) method in which cultured cells are fluorescently labeled to visualize subcellular structures (i.e., nucleus, nucleoli, endoplasmic reticulum, cytoskeleton, Golgi apparatus / plasma membrane and mitochondria) and to quantify morphological changes in response to chemicals or other perturbagens. HTPP is a high-throughput and cost-effective bioactivity screening method that detects effects associated with many different molecular mechanisms in an untargeted manner, enabling rapid in vitro hazard assessment for thousands of chemicals. Here, 1201 chemicals from the ToxCast library were screened in concentration-response up to ~100 μM in human U-2 OS cells using HTPP. A phenotype altering concentration (PAC) was estimated for chemicals active in the tested range. PACs tended to be higher than lower bound potency values estimated from a broad collection of targeted high-throughput assays, but lower than the threshold for cytotoxicity. In vitro to in vivo extrapolation (IVIVE) was used to estimate administered equivalent doses (AEDs) based on PACs for comparison to human exposure predictions. AEDs for 18/412 chemicals overlapped with predicted human exposures. Phenotypic profile information was also leveraged to identify putative mechanisms of action and group chemicals. Of 58 known nuclear receptor modulators, only glucocorticoids and retinoids produced characteristic profiles; and both receptor types are expressed in U-2 OS cells. Thirteen chemicals with profile similarity to glucocorticoids were tested in a secondary screen and one chemical, pyrene, was confirmed by an orthogonal gene expression assay as a novel putative GR modulating chemical. Most active chemicals demonstrated profiles not associated with a known mechanism-of-action. However, many structurally related chemicals produced similar profiles, with exceptions such as diniconazole, whose profile differed from other active conazoles. Overall, the present study demonstrates how HTPP can be applied in screening-level chemical assessments through a series of examples and brief case studies.

Selection of Optimal Cell Lines for High-Content Phenotypic Screening

High-content microscopy offers a scalable approach to screen against multiple targets in a single pass. Prior work has focused on methods to select "optimal" cellular readouts in microscopy screens. However, methods to select optimal cell line models have garnered much less attention. Here, we provide a roadmap for how to select the cell line or lines that are best suited to identify bioactive compounds and their mechanism of action (MOA). We test our approach on compounds targeting cancer-relevant pathways, ranking cell lines in two tasks: detecting compound activity ("phenoactivity") and grouping compounds with similar MOA by similar phenotype ("phenosimilarity"). Evaluating six cell lines across 3214 well-annotated compounds, we show that optimal cell line selection depends on both the task of interest (e.g., detecting phenoactivity vs inferring phenosimilarity) and distribution of MOAs within the compound library. Given a task of interest and a set of compounds, we provide a systematic framework for choosing optimal cell line(s). Our framework can be used to reduce the number of cell lines required to identify hits within a compound library and help accelerate the pace of early drug discovery.

Selection of optimal cell lines for high-content phenotypic screening

High-content microscopy offers a scalable approach to screen against multiple targets in a single pass. Prior work has focused on methods to select "optimal" cellular readouts in microscopy screens. However, methods to select optimal cell line models have garnered much less attention. Here, we provide a roadmap for how to select the cell line or lines that are best suited to identify bioactive compounds and their mechanism of action (MOA). We test our approach on compounds targeting cancer-relevant pathways, ranking cell lines in two tasks: detecting compound activity ("phenoactivity") and grouping compounds with similar MOA by similar phenotype ("phenosimilarity"). Evaluating six cell lines across 3214 well-annotated compounds, we show that optimal cell line selection depends on both the task of interest (e.g. detecting phenoactivity vs. inferring phenosimilarity) and distribution of MOAs within the compound library. Given a task of interest and set of compounds, we provide a systematic framework for choosing optimal cell line(s). Our framework can be used to reduce the number of cell lines required to identify hits within a compound library and help accelerate the pace of early drug discovery.

Cell segmentation and representation with shape priors

Cell segmentation is a fundamental problem of computational biology, for which convolutional neural networks yield the best results nowadays. This field is expanding rapidly, and in the recent years, shape-constrained segmentation models emerged as strong competitors to traditional, pixel-based segmentation methods for instance segmentation. These methods predict the parameters of the underlying shape model, so choosing the right shape representation is critical for the success of the segmentation. In this study, we introduce two new representation-based deep learning segmentation methods after a quantitative comparison of the most important shape descriptors in the literature. Our networks are based on Fourier coefficients and statistical shape models, both of which have proven to be reliable tools for cell shape modelling. Our results indicate that the methods are competitive alternatives to the most widely used baseline deep learning algorithms, especially when the number of parameters for the underlying shape model are low or the cells to be segmented have irregular morphologies.

Fisheye transformation enhances deep-learning-based single-cell phenotyping by including cellular microenvironment

Incorporating information about the surroundings can have a significant impact on successfully determining the class of an object. This is of particular interest when determining the phenotypes of cells, for example, in the context of high-throughput screens. We hypothesized that an ideal approach would consider the fully featured view of the cell of interest, include its neighboring microenvironment, and give lesser weight to cells that are far from the cell of interest. To satisfy these criteria, we present an approach with a transformation similar to those characteristic of fisheye cameras. Using this transformation with proper settings, we could significantly increase the accuracy of single-cell phenotyping, both in the case of cell culture and tissue-based microscopy images, and we present improved results on a dataset containing images of wild animals.

Applications of Neural Networks in Biomedical Data Analysis

Neural networks for deep-learning applications, also called artificial neural networks, are important tools in science and industry. While their widespread use was limited because of inadequate hardware in the past, their popularity increased dramatically starting in the early 2000s when it became possible to train increasingly large and complex networks. Today, deep learning is widely used in biomedicine from image analysis to diagnostics. This also includes special topics, such as forensics. In this review, we discuss the latest networks and how they work, with a focus on the analysis of biomedical data, particularly biomarkers in bioimage data. We provide a summary on numerous technical aspects, such as activation functions and frameworks. We also present a data analysis of publications about neural networks to provide a quantitative insight into the use of network types and the number of journals per year to determine the usage in different scientific fields.

Computational Methods for Single-Cell Imaging and Omics Data Integration

Integrating single cell omics and single cell imaging allows for a more effective characterisation of the underlying mechanisms that drive a phenotype at the tissue level, creating a comprehensive profile at the cellular level. Although the use of imaging data is well established in biomedical research, its primary application has been to observe phenotypes at the tissue or organ level, often using medical imaging techniques such as MRI, CT, and PET. These imaging technologies complement omics-based data in biomedical research because they are helpful for identifying associations between genotype and phenotype, along with functional changes occurring at the tissue level. Single cell imaging can act as an intermediary between these levels. Meanwhile new technologies continue to arrive that can be used to interrogate the genome of single cells and its related omics datasets. As these two areas, single cell imaging and single cell omics, each advance independently with the development of novel techniques, the opportunity to integrate these data types becomes more and more attractive. This review outlines some of the technologies and methods currently available for generating, processing, and analysing single-cell omics- and imaging data, and how they could be integrated to further our understanding of complex biological phenomena like ageing. We include an emphasis on machine learning algorithms because of their ability to identify complex patterns in large multidimensional data.

High-content analysis and Kinetic Image Cytometry identify toxicity and epigenetic effects of HIV antiretrovirals on human iPSC-neurons and primary neural precursor cells

INTRODUCTION

Despite viral suppression due to combination antiretroviral therapy (cART), HIV-associated neurocognitive disorders (HAND) continue to affect half of people with HIV, suggesting that certain antiretrovirals (ARVs) may contribute to HAND.

METHODS

We examined the effects of nucleoside/nucleotide reverse transcriptase inhibitors tenofovir disoproxil fumarate (TDF) and emtricitabine (FTC) and the integrase inhibitors dolutegravir (DTG) and elvitegravir (EVG) on viability, structure, and function of glutamatergic neurons (a subtype of CNS neuron involved in cognition) derived from human induced pluripotent stem cells (hiPSC-neurons), and primary human neural precursor cells (hNPCs), which are responsible for neurogenesis.

RESULTS

Using automated digital microscopy and image analysis (high content analysis, HCA), we found that DTG, EVG, and TDF decreased hiPSC-neuron viability, neurites, and synapses after 7 days of treatment. Analysis of hiPSC-neuron calcium activity using Kinetic Image Cytometry (KIC) demonstrated that DTG and EVG also decreased the frequency and magnitude of intracellular calcium transients. Longer ARV exposures and simultaneous exposure to multiple ARVs increased the magnitude of these neurotoxic effects. Using the Microscopic Imaging of Epigenetic Landscapes (MIEL) assay, we found that TDF decreased hNPC viability and changed the distribution of histone modifications that regulate chromatin packing, suggesting that TDF may reduce neuroprogenitor pools important for CNS development and maintenance of cognition in adults.

CONCLUSION

This study establishes human preclinical assays that can screen potential ARVs for CNS toxicity to develop safer cART regimens and HAND therapeutics.

Morphological profiling by means of the Cell Painting assay enables identification of tubulin-targeting compounds

In phenotypic compound discovery, conclusive identification of cellular targets and mode of action are often impaired by off-target binding. In particular, microtubules are frequently targeted in cellular assays. However, in vitro tubulin binding assays do not correctly reflect the cellular context, and conclusive high-throughput phenotypic assays monitoring tubulin binding are scarce, such that tubulin binding is rarely identified. We report that morphological profiling using the Cell Painting assay (CPA) can efficiently detect tubulin modulators in compound collections with a high throughput, including annotated reference compounds and unannotated compound classes with unrelated chemotypes and scaffolds. Small-molecule tubulin binders share similar CPA fingerprints, which enables prediction and experimental validation of microtubule-binding activity. Our findings suggest that CPA or a related morphological profiling approach will be an invaluable addition to small-molecule discovery programs in chemical biology and medicinal chemistry, enabling early identification of one of the most frequently observed off-target activities.

Deep learning enables confocal laser-scanning microscopy with enhanced resolution

Theoretical resolution enhancement of confocal laser-scanning microscopy (CLSM) is sacrificed for the best compromise between optical sectioning and the signal-to-noise ratio (SNR). The pixel reassignment reconstruction algorithm can improve the effective spatial resolution of CLSM to its theoretical limit. However, current implementations are not versatile and are time-consuming or technically complex. Here we present a parameter-free post-processing strategy for laser-scanning microscopy based on deep learning, which enables a spatial resolution enhancement by a factor of ∼1.3, compared to conventional CLSM. To speed up the training process for experimental data, transfer learning, combined with a hybrid dataset consisting of simulated synthetic and experimental images, is employed. The overall resolution and SNR improvement, validated by quantitative evaluation metrics, allowed us to correctly infer the fine structures of real experimental images.

Fully unsupervised deep mode of action learning for phenotyping high-content cellular images

MOTIVATION

The identification and discovery of phenotypes from high content screening (HCS) images is a challenging task. Earlier works use image analysis pipelines to extract biological features, supervised training methods or generate features with neural networks pretrained on non-cellular images. We introduce a novel unsupervised deep learning algorithm to cluster cellular images with similar Mode-of-Action (MOA) together using only the images' pixel intensity values as input. It corrects for batch effect during training. Importantly, our method does not require the extraction of cell candidates and works from the entire images directly.

RESULTS

The method achieves competitive results on the labelled subset of the BBBC021 dataset with an accuracy of 97.09% for correctly classifying the MOA by nearest neighbors matching. Importantly, we can train our approach on unannotated datasets. Therefore, our method can discover novel MOAs and annotate unlabelled compounds. The ability to train end-to-end on the full resolution images makes our method easy to apply and allows it to further distinguish treatments by their effect on proliferation.

AVAILABILITY

Our code is available at https://github.com/Novartis/UMM-Discovery.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Morphological profiling of human T and NK lymphocytes by high-content cell imaging

The immunological synapse is a complex structure that decodes stimulatory signals into adapted lymphocyte responses. It is a unique window to monitor lymphocyte activity because of development of systematic quantitative approaches. Here we demonstrate the applicability of high-content imaging to human T and natural killer (NK) cells and develop a pipeline for unbiased analysis of high-definition morphological profiles. Our approach reveals how distinct facets of actin cytoskeleton remodeling shape immunological synapse architecture and affect lytic granule positioning. Morphological profiling of CD8⁺ T cells from immunodeficient individuals allows discrimination of the roles of the ARP2/3 subunit ARPC1B and the ARP2/3 activator Wiskott-Aldrich syndrome protein (WASP) in immunological synapse assembly. Single-cell analysis further identifies uncoupling of lytic granules and F-actin radial distribution in ARPC1B-deficient lymphocytes. Our study provides a foundation for development of morphological profiling as a scalable approach to monitor primary lymphocyte responsiveness and to identify complex aspects of lymphocyte micro-architecture.

Image-based cell phenotyping with deep learning

A cell's phenotype is the culmination of several cellular processes through a complex network of molecular interactions that ultimately result in a unique morphological signature. Visual cell phenotyping is the characterization and quantification of these observable cellular traits in images. Recently, cellular phenotyping has undergone a massive overhaul in terms of scale, resolution, and throughput, which is attributable to advances across electronic, optical, and chemical technologies for imaging cells. Coupled with the rapid acceleration of deep learning-based computational tools, these advances have opened up new avenues for innovation across a wide variety of high-throughput cell biology applications. Here, we review applications wherein deep learning is powering the recognition, profiling, and prediction of visual phenotypes to answer important biological questions. As the complexity and scale of imaging assays increase, deep learning offers computational solutions to elucidate the details of previously unexplored cellular phenotypes.

countfitteR: efficient selection of count distributions to assess DNA damage

Background

DNA double-strand breaks can be counted as discrete foci by imaging techniques. In personalized medicine and pharmacology, the analysis of counting data is relevant for numerous applications, e.g., for cancer and aging research and the evaluation of drug efficacy. By default, it is assumed to follow the Poisson distribution. This assumption, however, may lead to biased results and faulty conclusions in datasets with excess zero values (zero-inflation), a variance larger than the mean (overdispersion), or both. In such cases, the assumption of a Poisson distribution would skew the estimation of mean and variance, and other models like the negative binomial (NB), zero-inflated Poisson or zero-inflated NB distributions should be employed. The model chosen has an influence on the parameter estimation (mean value and confidence interval). Yet the choice of the suitable distribution model is not trivial.

Methods

To support, simplify and objectify this process, we have developed the countfitteR software as an R package. We used a Bayesian approach for distribution model selection and the shiny web application framework for interactive data analysis.

Results

We show the application of our software based on examples of DNA double-strand break count data from phenotypic imaging by multiplex fluorescence microscopy. In analyzing numerous datasets of molecular pharmacological markers (phosphorylated histone H2AX and p53 binding protein), countfitteR demonstrated an equal or superior statistical performance compared to the usually employed two-step procedure, with an overall power of up to 98%. In addition, it still gave information in cases with no result at all from the two-step procedure. In our data sample we found that the NB distribution was the most frequent, with the Poisson distribution taking second place.

Conclusions

countfitteR can perform an automated distribution model selection and thus support the data analysis and lead to objective statistically verifiable estimated values. Originally designed for the analysis of foci in biomedical image data, countfitteR can be used in a variety of areas where non-Poisson distributed counting data is prevalent.

High-content phenotypic and pathway profiling to advance drug discovery in diseases of unmet need

Conventional thinking in modern drug discovery postulates that the design of highly selective molecules which act on a single disease-associated target will yield safer and more effective drugs. However, high clinical attrition rates and the lack of progress in developing new effective treatments for many important diseases of unmet therapeutic need challenge this hypothesis. This assumption also impinges upon the efficiency of target agnostic phenotypic drug discovery strategies, where early target deconvolution is seen as a critical step to progress phenotypic hits. In this review we provide an overview of how emerging phenotypic and pathway-profiling technologies integrate to deconvolute the mechanism-of-action of phenotypic hits. We propose that such in-depth mechanistic profiling may support more efficient phenotypic drug discovery strategies that are designed to more appropriately address complex heterogeneous diseases of unmet need.

Morphological profiling of small molecules

Profiling approaches such as gene expression or proteome profiling generate small-molecule bioactivity profiles that describe a perturbed cellular state in a rather unbiased manner and have become indispensable tools in the evaluation of bioactive small molecules. Automated imaging and image analysis can record morphological alterations that are induced by small molecules through the detection of hundreds of morphological features in high-throughput experiments. Thus, morphological profiling has gained growing attention in academia and the pharmaceutical industry as it enables detection of bioactivity in compound collections in a broader biological context in the early stages of compound development and the drug-discovery process. Profiling may be used successfully to predict mode of action or targets of unexplored compounds and to uncover unanticipated activity for already characterized small molecules. Here, we review the reported approaches to morphological profiling and the kind of bioactivity that can be detected so far and, thus, predicted.

Reproducible image handling and analysis

Image data are universal in life sciences research. Their proper handling is not. A significant proportion of image data in research papers show signs of mishandling that undermine their interpretation. We propose that a precise description of the image processing and analysis applied is required to address this problem. A new norm for reporting reproducible image analyses will diminish mishandling, as it will alert co-authors, referees, and journals to aberrant image data processing or, if published nonetheless, it will document it to the reader. To promote this norm, we discuss the effectiveness of this approach and give some step-by-step instructions for publishing reproducible image data processing and analysis workflows.

Image-based profiling for drug discovery: due for a machine-learning upgrade?

Image-based profiling is a maturing strategy by which the rich information present in biological images is reduced to a multidimensional profile, a collection of extracted image-based features. These profiles can be mined for relevant patterns, revealing unexpected biological activity that is useful for many steps in the drug discovery process. Such applications include identifying disease-associated screenable phenotypes, understanding disease mechanisms and predicting a drug's activity, toxicity or mechanism of action. Several of these applications have been recently validated and have moved into production mode within academia and the pharmaceutical industry. Some of these have yielded disappointing results in practice but are now of renewed interest due to improved machine-learning strategies that better leverage image-based information. Although challenges remain, novel computational technologies such as deep learning and single-cell methods that better capture the biological information in images hold promise for accelerating drug discovery.

Use of artificial intelligence to enhance phenotypic drug discovery

Research and development (R&D) productivity across the pharmaceutical industry has received close scrutiny over the past two decades, especially taking into consideration reports of attrition rates and the colossal cost for drug development. The respective merits of the two main drug discovery approaches, phenotypic and target based, have divided opinion across the research community, because each hold different advantages for identifying novel molecular entities with a successful path to the market. Nevertheless, both have low translatability in the clinic. Artificial intelligence (AI) and adoption of machine learning (ML) tools offer the promise of revolutionising drug development, and overcoming obstacles in the drug discovery pipeline. Here, we assess the potential of target-driven and phenotypic-based approaches and offer a holistic description of the current state of the field, from both a scientific and industry perspective. With the emerging partnerships between AI/ML and pharma still in their relative infancy, we investigate the potential and current limitations with a particular focus on phenotypic drug discovery. Finally, we emphasise the value of public-private partnerships (PPPs) and cross-disciplinary collaborations to foster innovation and facilitate efficient drug discovery programmes.

High Content Analysis Across Signaling Modulation Treatments for Subcellular Target Identification Reveals Heterogeneity in Cellular Response

Cellular phenotypes on bioactive compound treatment are a result of the downstream targets of the respective treatment. Here, a computational approach is taken for downstream subcellular target identification to understand the basis of the cellular response. This response is a readout of cellular phenotypes captured from cell-painting-based light microscopy images. The readouts are morphological profiles measured simultaneously from multiple cellular organelles. Cellular profiles generated from roughly 270 diverse treatments on bone cancer cell line form the high content screen used in this study. Phenotypic diversity across these treatments is demonstrated, depending on the image-based phenotypic profiles. Furthermore, the impact of the treatments on specific organelles and associated organelle sensitivities are determined. This revealed that endoplasmic reticulum has a higher likelihood of being targeted. Employing multivariate regression overall cellular response is predicted based on fewer organelle responses. This prediction model is validated against 1,000 new candidate compounds. Different compounds despite driving specific modulation outcomes elicit a varying effect on cellular integrity. Strikingly, this confirms that phenotypic responses are not conserved that enables quantification of signaling heterogeneity. Agonist-antagonist signaling pairs demonstrate switch of the targets in the cascades hinting toward evidence of signaling plasticity. Quantitative analysis of the screen has enabled the identification of these underlying signatures. Together, these image-based profiling approaches can be employed for target identification in drug and diseased states and understand the hallmark of cellular response.

Batch equalization with a generative adversarial network

Motivation

Advances in automation and imaging have made it possible to capture a large image dataset that spans multiple experimental batches of data. However, accurate biological comparison across the batches is challenged by batch-to-batch variation (i.e. batch effect) due to uncontrollable experimental noise (e.g. varying stain intensity or cell density). Previous approaches to minimize the batch effect have commonly focused on normalizing the low-dimensional image measurements such as an embedding generated by a neural network. However, normalization of the embedding could suffer from over-correction and alter true biological features (e.g. cell size) due to our limited ability to interpret the effect of the normalization on the embedding space. Although techniques like flat-field correction can be applied to normalize the image values directly, they are limited transformations that handle only simple artifacts due to batch effect.

Results

We present a neural network-based batch equalization method that can transfer images from one batch to another while preserving the biological phenotype. The equalization method is trained as a generative adversarial network (GAN), using the StarGAN architecture that has shown considerable ability in style transfer. After incorporating new objectives that disentangle batch effect from biological features, we show that the equalized images have less batch information and preserve the biological information. We also demonstrate that the same model training parameters can generalize to two dramatically different types of cells, indicating this approach could be broadly applicable.

Availability and implementation

https://github.com/tensorflow/gan/tree/master/tensorflow_gan/examples/stargan

Supplementary information

Supplementary data are available at Bioinformatics online.

TheCellVision.org: A Database for Visualizing and Mining High-Content Cell Imaging Projects

Advances in genome engineering and high throughput imaging technologies have enabled genome-scale screens of single cells for a variety of phenotypes, including subcellular morphology and protein localization. We constructed TheCellVision.org, a freely available and web-accessible image visualization and data browsing tool that serves as a central repository for fluorescence microscopy images and associated quantitative data produced by high-content screening experiments. Currently, TheCellVision.org hosts ∼575,590 images and associated analysis results from two published high-content screening (HCS) projects focused on the budding yeast Saccharomyces cerevisiae TheCellVision.org allows users to access, visualize and explore fluorescence microscopy images, and to search, compare, and extract data related to subcellular compartment morphology, protein abundance, and localization. Each dataset can be queried independently or as part of a search across multiple datasets using the advanced search option. The website also hosts computational tools associated with the available datasets, which can be applied to other projects and cell systems, a feature we demonstrate using published images of mammalian cells. Providing access to HCS data through websites such as TheCelllVision.org enables new discovery and independent re-analyses of imaging data.

ALICE: a hybrid AI paradigm with enhanced connectivity and cybersecurity for a serendipitous encounter with circulating hybrid cells

A fully automated and accurate assay of rare cell phenotypes in densely-packed fluorescently-labeled liquid biopsy images remains elusive. Methods: Employing a hybrid artificial intelligence (AI) paradigm that combines traditional rule-based morphological manipulations with modern statistical machine learning, we deployed a next generation software, ALICE (Automated Liquid Biopsy Cell Enumerator) to identify and enumerate minute amounts of tumor cell phenotypes bestrewed in massive populations of leukocytes. As a code designed for futurity, ALICE is armed with internet of things (IOT) connectivity to promote pedagogy and continuing education and also, an advanced cybersecurity system to safeguard against digital attacks from malicious data tampering. Results: By combining robust principal component analysis, random forest classifier and cubic support vector machine, ALICE was able to detect synthetic, anomalous and tampered input images with an average recall and precision of 0.840 and 0.752, respectively. In terms of phenotyping enumeration, ALICE was able to enumerate various circulating tumor cell (CTC) phenotypes with a reliability ranging from 0.725 (substantial agreement) to 0.961 (almost perfect) as compared to human analysts. Further, two subpopulations of circulating hybrid cells (CHCs) were serendipitously discovered and labeled as CHC-1 (DAPI+/CD45+/E-cadherin+/vimentin-) and CHC-2 (DAPI+ /CD45+/E-cadherin+/vimentin+) in the peripheral blood of pancreatic cancer patients. CHC-1 was found to correlate with nodal staging and was able to classify lymph node metastasis with a sensitivity of 0.615 (95% CI: 0.374-0.898) and specificity of 1.000 (95% CI: 1.000-1.000). Conclusion: This study presented a machine-learning-augmented rule-based hybrid AI algorithm with enhanced cybersecurity and connectivity for the automatic and flexibly-adapting enumeration of cellular liquid biopsies. ALICE has the potential to be used in a clinical setting for an accurate and reliable enumeration of CTC phenotypes.

Tissue dynamics spectroscopic imaging: functional imaging of heterogeneous cancer tissue

SIGNIFICANCE

Tumor heterogeneity poses a challenge for the chemotherapeutic treatment of cancer. Tissue dynamics spectroscopy captures dynamic contrast and can capture the response of living tissue to applied therapeutics, but the current analysis averages over the complicated spatial response of living biopsy samples.

AIM

To develop tissue dynamics spectroscopic imaging (TDSI) to map the heterogeneous spatial response of tumor tissue to anticancer drugs.

APPROACH

TDSI is applied to tumor spheroids grown from cell lines and to ex vivo living esophageal biopsy samples. Doppler fluctuation spectroscopy is performed on a voxel basis to extract spatial maps of biodynamic biomarkers. Functional images and bivariate spatial maps are produced using a bivariate color merge to represent the spatial distribution of pairs of signed drug-response biodynamic biomarkers.

RESULTS

We have mapped the spatial variability of drug responses within biopsies and have tracked sample-to-sample variability. Sample heterogeneity observed in the biodynamic maps is associated with histological heterogeneity observed using inverted selective-plane illumination microscopy.

CONCLUSION

We have demonstrated the utility of TDSI as a functional imaging method to measure tumor heterogeneity and its potential for use in drug-response profiling.

Comparison of Approaches for Determining Bioactivity Hits from High-Dimensional Profiling Data

Phenotypic profiling assays are untargeted screening assays that measure a large number (hundreds to thousands) of cellular features in response to a stimulus and often yield diverse and unanticipated profiles of phenotypic effects, leading to challenges in distinguishing active from inactive treatments. Here, we compare a variety of different strategies for hit identification in imaging-based phenotypic profiling assays using a previously published Cell Painting data set. Hit identification strategies based on multiconcentration analysis involve curve fitting at several levels of data aggregation (e.g., individual feature level, aggregation of similarly derived features into categories, and global modeling of all features) and on computed metrics (e.g., Euclidean and Mahalanobis distance metrics and eigenfeatures). Hit identification strategies based on single-concentration analysis included measurement of signal strength (e.g., total effect magnitude) and correlation of profiles among biological replicates. Modeling parameters for each approach were optimized to retain the ability to detect a reference chemical with subtle phenotypic effects while limiting the false-positive rate to 10%. The percentage of test chemicals identified as hits was highest for feature-level and category-based approaches, followed by global fitting, whereas signal strength and profile correlation approaches detected the fewest number of active hits at the fixed false-positive rate. Approaches involving fitting of distance metrics had the lowest likelihood for identifying high-potency false-positive hits that may be associated with assay noise. Most of the methods achieved a 100% hit rate for the reference chemical and high concordance for 82% of test chemicals, indicating that hit calls are robust across different analysis approaches.

Correction of refractive index mismatch-induced aberrations under radially polarized illumination by deep learning

Radially polarized field under strong focusing has emerged as a powerful manner for fluorescence microscopy. However, the refractive index (RI) mismatch-induced aberrations seriously degrade imaging performance, especially under high numerical aperture (NA). Traditional adaptive optics (AO) method is limited by its tedious procedure. Here, we present a computational strategy that uses artificial neural networks to correct the aberrations induced by RI mismatch. There are no requirements for expensive hardware and complicated wavefront sensing in our framework when the deep network training is completed. The structural similarity index (SSIM) criteria and spatial frequency spectrum analysis demonstrate that our deep-learning-based method has a better performance compared to the widely used Richardson-Lucy (RL) deconvolution method at different imaging depth on simulation data. Additionally, the generalization of our trained network model is tested on new types of samples that are not present in the training procedure to further evaluate the utility of the network, and the performance is also superior to RL deconvolution.

3D imaging of colorectal cancer organoids identifies responses to Tankyrase inhibitors

Aberrant activation of the Wnt signalling pathway is required for tumour initiation and survival in the majority of colorectal cancers. The development of inhibitors of Wnt signalling has been the focus of multiple drug discovery programs targeting colorectal cancer and other malignancies associated with aberrant pathway activation. However, progression of new clinical entities targeting the Wnt pathway has been slow. One challenge lies with the limited predictive power of 2D cancer cell lines because they fail to fully recapitulate intratumoural phenotypic heterogeneity. In particular, the relationship between 2D cancer cell biology and cancer stem cell function is poorly understood. By contrast, 3D tumour organoids provide a platform in which complex cell-cell interactions can be studied. However, complex 3D models provide a challenging platform for the quantitative analysis of drug responses of therapies that have differential effects on tumour cell subpopulations. Here, we generated tumour organoids from colorectal cancer patients and tested their responses to inhibitors of Tankyrase (TNKSi) which are known to modulate Wnt signalling. Using compounds with 3 orders of magnitude difference in cellular mechanistic potency together with image-based assays, we demonstrate that morphometric analyses can capture subtle alterations in organoid responses to Wnt inhibitors that are consistent with activity against a cancer stem cell subpopulation. Overall our study highlights the value of phenotypic readouts as a quantitative method to asses drug-induced effects in a relevant preclinical model.

High-Content Phenotypic Profiling in Esophageal Adenocarcinoma Identifies Selectively Active Pharmacological Classes of Drugs for Repurposing and Chemical Starting Points for Novel Drug Discovery

Esophageal adenocarcinoma (EAC) is a highly heterogeneous disease, dominated by large-scale genomic rearrangements and copy number alterations. Such characteristics have hampered conventional target-directed drug discovery and personalized medicine strategies, contributing to poor outcomes for patients. We describe the application of a high-content Cell Painting assay to profile the phenotypic response of 19,555 compounds across a panel of six EAC cell lines and two tissue-matched control lines. We built an automated high-content image analysis pipeline to identify compounds that selectively modified the phenotype of EAC cell lines. We further trained a machine-learning model to predict the mechanism of action of EAC selective compounds using phenotypic fingerprints from a library of reference compounds. We identified a number of phenotypic clusters enriched with similar pharmacological classes, including methotrexate and three other antimetabolites that are highly selective for EAC cell lines. We further identify a small number of hits from our diverse chemical library that show potent and selective activity for EAC cell lines and that do not cluster with the reference library of compounds, indicating they may be selectively targeting novel esophageal cancer biology. Overall, our results demonstrate that our EAC phenotypic screening platform can identify existing pharmacologic classes and novel compounds with selective activity for EAC cell phenotypes.

Evaluation of methods for generative modeling of cell and nuclear shape

MOTIVATION

Cell shape provides both geometry for, and a reflection of, cell function. Numerous methods for describing and modeling cell shape have been described, but previous evaluation of these methods in terms of the accuracy of generative models has been limited.

RESULTS

Here we compare traditional methods and deep autoencoders to build generative models for cell shapes in terms of the accuracy with which shapes can be reconstructed from models. We evaluated the methods on different collections of 2D and 3D cell images, and found that none of the methods gave accurate reconstructions using low dimensional encodings. As expected, much higher accuracies were observed using high dimensional encodings, with outline-based methods significantly outperforming image-based autoencoders. The latter tended to encode all cells as having smooth shapes, even for high dimensions. For complex 3D cell shapes, we developed a significant improvement of a method based on the spherical harmonic transform that performs significantly better than other methods. We obtained similar results for the joint modeling of cell and nuclear shape. Finally, we evaluated the modeling of shape dynamics by interpolation in the shape space. We found that our modified method provided lower deformation energies along linear interpolation paths than other methods. This allows practical shape evolution in high dimensional shape spaces. We conclude that our improved spherical harmonic based methods are preferable for cell and nuclear shape modeling, providing better representations, higher computational efficiency and requiring fewer training images than deep learning methods.

AVAILABILITY AND IMPLEMENTATION

All software and data is available at http://murphylab.cbd.cmu.edu/software.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Test-time augmentation for deep learning-based cell segmentation on microscopy images

Recent advancements in deep learning have revolutionized the way microscopy images of cells are processed. Deep learning network architectures have a large number of parameters, thus, in order to reach high accuracy, they require a massive amount of annotated data. A common way of improving accuracy builds on the artificial increase of the training set by using different augmentation techniques. A less common way relies on test-time augmentation (TTA) which yields transformed versions of the image for prediction and the results are merged. In this paper we describe how we have incorporated the test-time argumentation prediction method into two major segmentation approaches utilized in the single-cell analysis of microscopy images. These approaches are semantic segmentation based on the U-Net, and instance segmentation based on the Mask R-CNN models. Our findings show that even if only simple test-time augmentations (such as rotation or flipping and proper merging methods) are applied, TTA can significantly improve prediction accuracy. We have utilized images of tissue and cell cultures from the Data Science Bowl (DSB) 2018 nuclei segmentation competition and other sources. Additionally, boosting the highest-scoring method of the DSB with TTA, we could further improve prediction accuracy, and our method has reached an ever-best score at the DSB.

Correcting nuisance variation using Wasserstein distance

Profiling cellular phenotypes from microscopic imaging can provide meaningful biological information resulting from various factors affecting the cells. One motivating application is drug development: morphological cell features can be captured from images, from which similarities between different drug compounds applied at different doses can be quantified. The general approach is to find a function mapping the images to an embedding space of manageable dimensionality whose geometry captures relevant features of the input images. An important known issue for such methods is separating relevant biological signal from nuisance variation. For example, the embedding vectors tend to be more correlated for cells that were cultured and imaged during the same week than for those from different weeks, despite having identical drug compounds applied in both cases. In this case, the particular batch in which a set of experiments were conducted constitutes the domain of the data; an ideal set of image embeddings should contain only the relevant biological information (e.g., drug effects). We develop a general framework for adjusting the image embeddings in order to “forget” domain-specific information while preserving relevant biological information. To achieve this, we minimize a loss function based on distances between marginal distributions (such as the Wasserstein distance) of embeddings across domains for each replicated treatment. For the dataset we present results with, the only replicated treatment happens to be the negative control treatment, for which we do not expect any treatment-induced cell morphology changes. We find that for our transformed embeddings (i) the underlying geometric structure is not only preserved but the embeddings also carry improved biological signal; and (ii) less domain-specific information is present.

A Synergistic Anticancer FAK and HDAC Inhibitor Combination Discovered by a Novel Chemical-Genetic High-Content Phenotypic Screen

We mutated the focal adhesion kinase (FAK) catalytic domain to inhibit binding of the chaperone Cdc37 and ATP, mimicking the actions of a FAK kinase inhibitor. We reexpressed mutant and wild-type FAK in squamous cell carcinoma (SCC) cells from which endogenous FAK had been deleted, genetically fixing one axis of a FAK inhibitor combination high-content phenotypic screen to discover drugs that may synergize with FAK inhibitors. Histone deacetylase (HDAC) inhibitors represented the major class of compounds that potently induced multiparametric phenotypic changes when FAK was rendered kinase-defective or inhibited pharmacologically in SCC cells. Combined FAK and HDAC inhibitors arrest proliferation and induce apoptosis in a subset of cancer cell lines in vitro and efficiently inhibit their growth as tumors in vivo Mechanistically, HDAC inhibitors potentiate inhibitor-induced FAK inactivation and impair FAK-associated nuclear YAP in sensitive cancer cell lines. Here, we report the discovery of a new, clinically actionable, synergistic combination between FAK and HDAC inhibitors.

High content phenotypic screening identifies serotonin receptor modulators with selective activity upon breast cancer cell cycle and cytokine signaling pathways

Heterogeneity in disease mechanisms between genetically distinct patients contributes to high attrition rates in late stage clinical drug development. New personalized medicine strategies aim to identify predictive biomarkers which stratify patients most likely to respond to a particular therapy. However, for complex multifactorial diseases not characterized by a single genetic driver, empirical approaches to identifying predictive biomarkers and the most promising therapies for personalized medicine are required. In vitro pharmacogenomics seeks to correlate in vitro drug sensitivity testing across panels of genetically distinct cell models with genomic, gene expression or proteomic data to identify predictive biomarkers of drug response. However, the vast majority of in vitro pharmacogenomic studies performed to date are limited to dose-response screening upon a single viability assay endpoint. In this article we describe the application of multiparametric high content phenotypic screening and the theta comparative cell scoring method to quantify and rank compound hits, screened at a single concentration, which induce a broad variety of divergent phenotypic responses between distinct breast cancer cell lines. High content screening followed by transcriptomic pathway analysis identified serotonin receptor modulators which display selective activity upon breast cancer cell cycle and cytokine signaling pathways correlating with inhibition of cell growth and survival. These methods describe a new evidence-led approach to rapidly identify compounds which display distinct response between different cell types. The results presented also warrant further investigation of the selective activity of serotonin receptor modulators upon breast cancer cell growth and survival as a potential drug repurposing opportunity.

Improving drug discovery using image-based multiparametric analysis of the epigenetic landscape

High-content phenotypic screening has become the approach of choice for drug discovery due to its ability to extract drug-specific multi-layered data. In the field of epigenetics, such screening methods have suffered from a lack of tools sensitive to selective epigenetic perturbations. Here we describe a novel approach, Microscopic Imaging of Epigenetic Landscapes (MIEL), which captures the nuclear staining patterns of epigenetic marks and employs machine learning to accurately distinguish between such patterns. We validated the MIEL platform across multiple cells lines and using dose-response curves, to insure the fidelity and robustness of this approach for high content high throughput drug discovery. Focusing on noncytotoxic glioblastoma treatments, we demonstrated that MIEL can identify and classify epigenetically active drugs. Furthermore, we show MIEL was able to accurately rank candidate drugs by their ability to produce desired epigenetic alterations consistent with increased sensitivity to chemotherapeutic agents or with induction of glioblastoma differentiation.

Multidimensional profiling of drug-treated cells by Imaging Mass Cytometry

In pharmaceutical research, high‐content screening is an integral part of lead candidate development. Measuring drug response in vitro by examining over 40 parameters, including biomarkers, signaling molecules, cell morphological changes, proliferation indices, and toxicity in a single sample, could significantly enhance discovery of new therapeutics. As a proof of concept, we present here a workflow for multidimensional Imaging Mass Cytometry™ (IMC™) and data processing with open source computational tools. CellProfiler was used to identify single cells through establishing cellular boundaries, followed by histoCAT™ (histology topography cytometry analysis toolbox) for extracting single‐cell quantitative information visualized as t‐SNE plots and heatmaps. Human breast cancer‐derived cell lines SKBR3, HCC1143, and MCF‐7 were screened for expression of cellular markers to generate digital images with a resolution comparable to conventional fluorescence microscopy. Predicted pharmacodynamic effects were measured in MCF‐7 cells dosed with three target‐specific compounds: growth stimulatory EGF, microtubule depolymerization agent nocodazole, and genotoxic chemotherapeutic drug etoposide. We show strong pairwise correlation between nuclear markers pHistone3^S28, Ki‐67, and p4E‐BP1^T37/T46 in classified mitotic cells and anticorrelation with cell surface markers. Our study demonstrates that IMC data expand the number of measured parameters in single cells and brings higher‐dimension analysis to the field of cell‐based screening in early lead compound discovery.

Machine Learning with Optical Phase Signatures for Phenotypic Profiling of Cell Lines

Robust and reproducible profiling of cell lines is essential for phenotypic screening assays. The goals of this study were to determine robust and reproducible optical phase signatures of cell lines for classification with machine learning and to correlate optical phase parameters to motile behavior. Digital holographic microscopy (DHM) reconstructed phase maps of cells from two pairs of cancer and non-cancer cell lines. Seventeen image parameters were extracted from each cell's phase map, used for linear support vector machine learning, and correlated to scratch wound closure and Boyden chamber chemotaxis. The classification accuracy was between 90% and 100% for the six pairwise cell line comparisons. Several phase parameters correlated with wound closure rate and chemotaxis across the four cell lines. The level of cell confluence in culture affected phase parameters in all cell lines tested. Results indicate that optical phase features of cell lines are a robust set of quantitative data of potential utility for phenotypic screening and prediction of motile behavior. © 2019 International Society for Advancement of Cytometry.

Chemical biology approaches targeting the actin cytoskeleton through phenotypic screening

The actin cytoskeleton is dysregulated in cancer, yet this critical cellular machinery has not translated as a druggable clinical target due to cardio-toxic side-effects. Many actin regulators are also considered undruggable, being structural proteins lacking clear functional sites suitable for targeted drug design. In this review, we discuss opportunities and challenges associated with drugging the actin cytoskeleton through its structural regulators, taking tropomyosins as a target example. In particular, we highlight emerging data acquisition and analysis trends driving phenotypic, imaging-based compound screening. Finally, we consider how the confluence of these trends is now bringing functionally integral machineries such as the actin cytoskeleton, and associated structural regulatory proteins, into an expanded repertoire of druggable targets with previously unexploited clinical potential.

Unbiased Phenotype Detection Using Negative Controls

Phenotypic screens using automated microscopy allow comprehensive measurement of the effects of compounds on cells due to the number of markers that can be scored and the richness of the parameters that can be extracted. The high dimensionality of the data is both a rich source of information and a source of noise that might hide information. Many methods have been proposed to deal with this complex data in order to reduce the complexity and identify interesting phenotypes. Nevertheless, the majority of laboratories still only use one or two parameters in their analysis, likely due to the computational challenges of carrying out a more sophisticated analysis. Here, we present a novel method that allows discovering new, previously unknown phenotypes based on negative controls only. The method is compared with L1-norm regularization, a standard method to obtain a sparse matrix. The analytical pipeline is implemented in the open-source software KNIME, allowing the implementation of the method in many laboratories, even ones without advanced computing knowledge.

Evaluation of Machine Learning Classifiers to Predict Compound Mechanism of Action When Transferred across Distinct Cell Lines

Multiparametric high-content imaging assays have become established to classify cell phenotypes from functional genomic and small-molecule library screening assays. Several groups have implemented machine learning classifiers to predict the mechanism of action of phenotypic hit compounds by comparing the similarity of their high-content phenotypic profiles with a reference library of well-annotated compounds. However, the majority of such examples are restricted to a single cell type often selected because of its suitability for simple image analysis and intuitive segmentation of morphological features. The aim of the current study was to evaluate and compare the performance of a classic ensemble-based tree classifier trained on extracted morphological features and a deep learning classifier using convolutional neural networks (CNNs) trained directly on images from the same dataset to predict compound mechanism of action across a morphologically and genetically distinct cell panel. Our results demonstrate that application of a CNN classifier delivers equivalent accuracy compared with an ensemble-based tree classifier at compound mechanism of action prediction within cell lines. However, our CNN analysis performs worse than an ensemble-based tree classifier when trained on multiple cell lines at predicting compound mechanism of action on an unseen cell line.

Transfer Learning with Deep Convolutional Neural Networks for Classifying Cellular Morphological Changes

The quantification and identification of cellular phenotypes from high-content microscopy images has proven to be very useful for understanding biological activity in response to different drug treatments. The traditional approach has been to use classical image analysis to quantify changes in cell morphology, which requires several nontrivial and independent analysis steps. Recently, convolutional neural networks have emerged as a compelling alternative, offering good predictive performance and the possibility to replace traditional workflows with a single network architecture. In this study, we applied the pretrained deep convolutional neural networks ResNet50, InceptionV3, and InceptionResnetV2 to predict cell mechanisms of action in response to chemical perturbations for two cell profiling datasets from the Broad Bioimage Benchmark Collection. These networks were pretrained on ImageNet, enabling much quicker model training. We obtain higher predictive accuracy than previously reported, between 95% and 97%. The ability to quickly and accurately distinguish between different cell morphologies from a scarce amount of labeled data illustrates the combined benefit of transfer learning and deep convolutional neural networks for interrogating cell-based images.

Machine learning and image-based profiling in drug discovery

The increase in imaging throughput, new analytical frameworks and high-performance computational resources open new avenues for data-rich phenotypic profiling of small molecules in drug discovery. Image-based profiling assays assessing single-cell phenotypes have been used to explore mechanisms of action, target efficacy and toxicity of small molecules. Technological advances to generate large data sets together with new machine learning approaches for the analysis of high-dimensional profiling data create opportunities to improve many steps in drug discovery. In this review, we will discuss how recent studies applied machine learning approaches in functional profiling workflows with a focus on chemical genetics. While their utility in image-based screening and profiling is predictably evident, examples of novel insights beyond the status quo based on the applications of machine learning approaches are just beginning to emerge. To enable discoveries, future studies also need to develop methodologies that lower the entry barriers to high-throughput profiling experiments by streamlining image-based profiling assays and providing applications for advanced learning technologies such as easy to deploy deep neural networks.

Environmental properties of cells improve machine learning-based phenotype recognition accuracy

To answer major questions of cell biology, it is often essential to understand the complex phenotypic composition of cellular systems precisely. Modern automated microscopes produce vast amounts of images routinely, making manual analysis nearly impossible. Due to their efficiency, machine learning-based analysis software have become essential tools to perform single-cell-level phenotypic analysis of large imaging datasets. However, an important limitation of such methods is that they do not use the information gained from the cellular micro- and macroenvironment: the algorithmic decision is based solely on the local properties of the cell of interest. Here, we present how various features from the surrounding environment contribute to identifying a cell and how such additional information can improve single-cell-level phenotypic image analysis. The proposed methodology was tested for different sizes of Euclidean and nearest neighbour-based cellular environments both on tissue sections and cell cultures. Our experimental data verify that the surrounding area of a cell largely determines its entity. This effect was found to be especially strong for established tissues, while it was somewhat weaker in the case of cell cultures. Our analysis shows that combining local cellular features with the properties of the cell’s neighbourhood significantly improves the accuracy of machine learning-based phenotyping.

Weakly Supervised Learning of Single-Cell Feature Embeddings

We study the problem of learning representations for single cells in microscopy images to discover biological relationships between their experimental conditions. Many new applications in drug discovery and functional genomics require capturing the morphology of individual cells as comprehensively as possible. Deep convolutional neural networks (CNNs) can learn powerful visual representations, but require ground truth for training; this is rarely available in biomedical profiling experiments. While we do not know which experimental treatments produce cells that look alike, we do know that cells exposed to the same experimental treatment should generally look similar. Thus, we explore training CNNs using a weakly supervised approach that uses this information for feature learning. In addition, the training stage is regularized to control for unwanted variations using mixup or RNNs. We conduct experiments on two different datasets; the proposed approach yields single-cell embeddings that are more accurate than the widely adopted classical features, and are competitive with previously proposed transfer learning approaches.