High-content screening (HCS) and the technology to do it faster, on more compounds in a shorter period of time, and to generate quantitative, multiparametric data took center stage at CHI’s “High Content East” meeting held in Boston last month. Presenters described how they are implementing enhanced screening systems, image-analysis methods, and data-management strategies to achieve daily HCS runs on tens of thousands of wells and screening campaigns totaling 200,000 to 3 million wells.
High throughput HCS—albeit not yet reaching the numbers common for conventional high-throughput screening (HTS) and with lingering limitations and challenges related to live-cell imaging over time—is making its mark and being used to probe the biological basis of disease and to detect even subtle phenotypic changes in response to experimental compounds.
Determining whether a cell looks like a cancer cell, for example, typically requires being able to detect subtle morphological changes, such as small alterations in size or structure, changes in the connections a cell makes with neighboring cells, or variations in the texture of staining. These have, historically, been mainly qualitative parameters detected by studying and comparing images of cells.
In her talk at the conference, Anne Carpenter, Ph.D., director of the imaging platform at the Broad Institute of Harvard University and MIT, presented her group’s work using HCS and image analysis to quantify difficult phenotypes and differentiate disease states such as leukemia.
Not only do HCS systems and image-analysis software automate the screening process, enabling theanalysis of many more cells in less time and increasing the chances of detecting even small numbers of altered cells, they can also utilize algorithms that evaluate defined combinations of parameters in a quantifiable manner and apply techniques to distinguish between clumping or closely juxtaposed cells. Relying on computer-based image analysis also standardizes the process, eliminating factors such as variability in human expertise and experience, consistency, and fatigue.
Dr. Carpenter’s group uses machine-learning methods to train image-analysis software to identify subtle phenotypic changes. Biologists work with the software in an iterative fashion in a process called supervised machine learning. They teach and correct the computers on a series of test images, refining the system’s knowledge base in a process that typically takes less than a day. The group developed the algorithms used by the biologists and has made them available as open-source software.
A recent paper published in PNAS by T. R. Jones, et al., documents the use of a trained image-analysis system to discriminate 15 different cellular phenotypes. Other projects involve teaching the software to discriminate leukemic from normal cells, to identify liver cells that are growing normally in culture—to aid in the development of physiologic models of liver function for use in drug testing—and training computers to detect subtle changes that signal the initiation of cell division for studying cell-cycle regulation in cancer.
Neil Carragher, Ph.D., senior scientist in the advanced science and technology laboratory at AstraZeneca, described how the company is applying high-content and live-cell imaging techniques and integrating the results with data derived from in vivo imaging and proteomic studies to improve clinical predictability.
Dr. Carragher’s group combines the results of high-content in vitro and in vivo assays to generate mechanistic information about phenotypic responses on candidate therapeutic compounds. The goal is to create a multiparametric fingerprint of a phenotype from images generated by HCS and to use this knowledge to enhance predictions of efficacy and toxicity early in drug discovery and reduce attrition later in development.
The phenotypic signatures are based on measurements of approximately 150 different parameters per cell for each assay. Data from multiple assays is collated for every test compound and compared with data obtained using well-characterized reference compounds to generate mechanistic hypotheses.
Only recently has open-source and commercial software become available “that allows you to quantitate more complex phenotypes, subtle changes, and heterogeneous responses from images,” Dr. Carragher said.
His group is employing two main approaches—each with different advantages and limitations. The first strategy relies on Definiens’ Cognition Network Technology™ software that allows users to develop algorithms that capture, computationally, what researchers can see visually. “It is very much context-based” and identifies objects based on how they are related to others in the image, rather than as individual pixels, explained Dr. Carragher. The in-house algorithm-development process depends on iterative programming steps. The other approach involves machine-learning tools using software such as the CellProfiler developed at the Broad Institute.
Redirecting Approved Drugs
Identifying new applications for FDA-approved drugs using HCS and image-based systems biology is the focus of work being done by Stephen Wong, Ph.D., founding director of the bioinformatics and biomedical engineering program and the cellular and tissue microscopy core at the Methodist Hospital Research Institute and professor of radiology and neurosciences at Weill Cornell Medical College.
Dr. Wong gave examples of screening campaigns to decipher targets in the pathways responsible for the metastasis of breast cancer to the brain in his talk. He specifically described the computational tools his group is developing for high-content and network analysis, and the animal-imaging techniques being used to evaluate combinations of small molecule chemotherapeutic agents for their ability to cross the blood-brain barrier and to have an effect against central nervous system metastases in breast cancer.
Dr. Wong’s group has also developed a series of quantitative image-analysis tools, including zebrafish image quantifier (ZFIQ), as well as software for studying neuronal spines (NeuronIQ), neurites (Neurite IQ), and time-lapse mitotic events in cells (DCellIQ). Dr. Wong’s HCS/systems biology research is funded by the NCI, NIA, and NLM.
Because the compounds being studied are already approved drugs, Phase I trials are not needed. The quantitative data generated from HCS provides the evidence necessary for moving into Phase II studies, shortening the drug-development cycle to a year or less.
The types of studies essential to Dr. Wong’s efforts, such as assays to monitor cell-cycle regulation or dendritic spine dynamics, require time-lapse, live-cell imaging. Looking at fixed cells provides only an artificial snapshot of where cells are at a particular point in time, explained Dr. Wong. “We want to look at a 384-well plate of continuously growing cells over five to six days,” he said, and in his view none of the instrument manufacturers competing in the HCS market has yet to provide a robust, incubator-based, environmentally controlled system that can achieve this.
Vendors have tended to view HCS as just another type of high-throughput screening, but live-cell imaging done in as natural an environment as possible has quite different requirements, contended Dr. Wong.
“Vendors are going in the wrong direction. The power of HCS is in the ability to visualize things in action and to extract lots more quantitative information from the images. If you, instead, retrofit HCS to HTS, you are losing its advantages,” such as the ability to see cells or spines change over time, to visualize cell-cell interactions, and to sync cell populations and study cell-cycle events in time-lapse, said Dr. Wong.
In any experiment, “if you generate enough data you will get hits, but how many will be real hits versus false positives?” asked Dr. Wong. “We need to push the quality upfront on the biology side” and screen out, earlier in the discovery process, compounds that are destined to fail.
Researchers at Pfizer are using HCS to study the genetic variation and physiologic interactions that underlie hepatic insulin resistance in type 2 diabetes and the prediabetic state. Diabetes is a complex, multigenic disease, and while advances in genomic and SNP-based technologies have led to the identification of at least 30 genes that contribute to the diabetic phenotype, much work remains to understand their role in cell biology and disease and how they interact.
“If you are careful about the cell models you choose, you can use HCS to characterize these genes and monitor their effects on biochemical pathways,” said Steven Haney, Ph.D., associate fellow in biological profiling at Pfizer’s biotherapeutics and bioinnovation center. The company has invested heavily in developing cell models that are representative of human physiology, including hepatocytes that faithfully mimic liver function when grown in culture.
The other main aspect of this research effort involves identifying changes that affect the diabetic phenotype, specifically glucose storage and utilization pathways, and distinguishing between effects that involve the insulin-signaling pathway from more general phenomena related to activation of toxicologic or stress pathways.
“HCS can alert us to things we don’t necessarily know to look for, in a mechanism-independent way,” said Dr. Haney. “The increasing throughput of HCS allows us to look at a lot of cells and determine whether subtle phenotypic changes are significant or spurious.”