This talk will peek under the hood to show how we combine deep learning models with biology lab automation at Recursion Pharmaceuticals. Every week, we generate millions of microscopy images of human cells on our high throughput screening platform as part of our search for new drug treatments. Here, I’ll describe how we use machine learning analytics to make sense of all of this complex data -  that is, to identify promising drug compounds while simultaneously avoiding high risk ones. As part of our aim to ‘industrialize drug discovery’, we have developed a single, large-volume platform approach that can be applied across large numbers of diseases. I will show how running such a high sophisticated experiment pipeline relies on a close integration of biology, engineering and data science product development: Engineering engineering infrastructure that can handle petabytes of data; Experimental biology setups that are optimized best enable downstream machine learning tasks; deep learning models that work not just on small, pre-defined datasets, but on large quantities of future data. Cumulatively, you will see a vision of medical discovery at an industrial scale.

Image-based high-content screening (HCS) involving cancer cell lines grown in conventional two-dimensional (2D) cell culture, has provided a cornerstone assay for searching new agents to treat pancreatic ductal adenocarcinoma (PDAC). However, since drug response in tumors in vivo is mediated by a myriad of microenvironmental factors (e.g tumor cytoarchitecture and cellular heterogeneity) the search for advanced cellular systems that can recapitulate more closely the tumor context have become a priority. Tumor spheroids consist in clusters obtained from a single cell population (e.g. 2D cell lines), whereas primary organoids can be established as self-organizing complex three-dimensional (3D) structures from cells of tumor resections from patients. All these tumor in-a-dish models better express chemoresistance mechanisms observed in cancers in vivo, and for this reason, they have been incorporated recently into drug HCS pipelines. The challenge now is to systematically visualize and quantitatively analyze chemoresistance-triggering pathways at the subcellular level, for which it will be necessary to engage image-based phenotypic profiling in those ‘tumor avatars’. We have explored the possibility of devising a HCS platform capable to map the spatial distribution of marker proteins to either cells or organelles within the multicellular structures. Here I present a 3D-HCS platform based on miniaturization of 3D culture in optical-bottom multi-well plates, endowed with a HCA-workflow that is effective at capturing cellular and multicellular morphological features in populations of PDAC spheroids and organoids. Our platform provides a perspective for application of this platform for a dissection of drug penetrance determinants.

The inner nuclear membrane of eukaryotic cells defines many aspects of gene expression and chromatin organization.  It is also therapeutically important as several diseases, termed “laminopathies” are related to the function of the inner nuclear membrane.  Nevertheless, the intimate connection with the outer nuclear membrane and the endoplasmic reticulum make it difficult to biochemically probe the proteome of the inner nuclear membrane without contamination.  We have leveraged the power of yeast genomics and split-GFP complementation to highlight only the inner nuclear membrane facing epitopes and performed a candidate screen of ~1000 genes in living yeast cells.  To our surprise, ~400 genes showed significant localization at the inner nuclear membrane.  A counter-screen with degradation pathway mutants revealed that the INMAD pathway and Asi1 is the dominant influence on the INM proteome.  In addition, we have performed low throughput follow up studies of protein-protein interactions unique to this compartment via fluorescence cross-correlation spectroscopy.  These have revealed unique molecular interactions in this compartment that can provide insight into unique roles of traditional ER proteins at the inner nuclear membrane.

The recent development of highly multiplexed methods for imaging tissues from patients and mouse models promises to fundamentally advance research in tissue biology and disease. These new technologies enable the study of cell states, cell-cell interactions, and tissue architecture in normal and disease conditions. A common application of tissue imaging in oncology is identifying and enumerating immune cell types and mapping their locations relative to tumor and stromal cells. Such spatially-resolved data is pertinent to understanding the mechanisms of action of immunotherapies. These tools however can be applied to investigate a wide range of topics including stress responses and cell cycle regulation. The ability to extract high-dimensional information from single cells within their native context provides a distinct opportunity to understand complex cell biologic principles underlying disease pathogenesis.

Image-based approaches to single-cell transcriptomics are emerging as powerful complements to single-cell RNA sequencing, in part, because these techniques preserve the native spatial context of RNAs within cells and tissues. I will describe multiplexed error robust single-molecule fluorescence in situ hybridization (MERFISH)—a technique capable of imaging thousands of different RNAs simultaneously in fixed cells—and its use for the discovery and mapping of cell types within intact tissues.

Imaging offers a means to draw upon the growing body of high-throughput molecular data to better understand the underlying cellular and molecular mechanisms of embryonic development; however, it is challenged by tradeoffs between speed, resolution, field of view and the photon budget. We are advancing this tradeoff by constructing two-photon light-sheet microscopes, combining the deep penetration of two-photon microscopy and the speed of light sheet microscopy, permitting 4D cell and molecular imaging with sufficient speed and resolution to generate unambiguous tracing of cells and signals in intact systems. To increase the 5 th Dimensions, we are refining a new generation of multispectral image analysis tools that exceed the performance of our previous work on Linear Unmixing by orders of magnitude in speed, error propagation and accuracy. These new analysis tools permit rapid and unambiguous analyses of multiplex-labeled specimens. Finally, to move to faster volumetric imaging, we have combined light field and light sheet approaches, offering the signal to noise needed to image thousands of neurons with high fidelity. Combined, these tools offer the multi-dimensional imaging required to follow key events in as they take place and allow us to use variance as an experimental tool rather than feeling its effects as a limitation.

Abnormal cellular metabolism is a hallmark of many diseases, yet there is an absence of quantitative methods to dynamically image metabolism with cellular-level resolution. Optical metabolic imaging (OMI) quantifies the fluorescence intensities and lifetimes of the metabolic co-enzymes NAD(P)H and FAD using two-photon microscopy. OMI is a label-free, high-resolution, quantitative tool for monitoring cellular metabolism within intact samples. OMI has been applied to 3D primary tumor organoids derived from patients to rapidly test multi-drug response. This platform has been validated in mouse models of breast and pancreas cancer, and feasibility has been tested in human tumors with chemotherapies, targeted therapies, radiation, and experimental drugs. The cellular-level assessment of OMI allows for sub-populations of cells with varying response to drug treatment to be tracked over time, to monitor therapeutic effect in resistant cell populations. OMI has recently been extended to image immune cell activation and function. These metabolic imaging tools have significant implications for rapid cellular-level assessment of metabolic response to drug treatment within engineered tissues, which could impact drug development and clinical treatment planning.

Immunotherapy such as immune checkpoint inhibitor antibodies have revolutionized the treatments for hard-to-treat cancers, with durable responses observed in clinics. However, the overall response is observed only in a small subset of patients. Also, immunotherapy induces delayed on-set of responses and novel patterns of the anti-tumor response which makes it challenging to identify patient responders and non-responders early on, often leading to undertreatment or overtreatment. To address these challenges, we engineered a ‘stimuli-responsive nanomaterial’ (SRN) that can not only deliver an immune checkpoint inhibitor to the tumor but also report back on its efficacy in real time. We rationalized that this could be achieved by a novel two-staged stimuli-responsive polymeric nanomaterial with well-defined ratio of an immunotherapy drug and a drug-function activatable imaging agent. To accomplish this, we engineered a stimuli-responsive nanomaterial which comprises of three building blocks: an immunotherapy drug (anti-PDL1 antibody), an enzyme activatable imaging agent and a polymeric backbone that holds both elements together. In preliminary studies, we observed that anti-PD-L1 antibody conjugated SRNs inhibited PD1-PDL1 interactions efficiently and induced T-cell mediated cancer cell apoptosis that can be imaged using activatable imaging probe. SRNs not only enabled real-time immunotherapy response imaging in tumor bearing mice but also distinguished between highly responsive and partially responsive tumors. Furthermore, increasing doses resulted in better response and enhanced sensitivity in partially responsive tumors. This study shows that the treatment with SRNs induced a potent anti-tumor immune response that can be directly imaged and is more effective in imaging response than current agents.

The Allen Institute for Cell Science is generating a state space of stem cell signatures. The goal is to understand cell organization, identify cell states, and elucidate how cells transition from state to state. We are doing this by conjoining high replicate 3D live cell imaging of cell lines gene-edited with GFP tagged proteins, single cell RNAseq, computational analyses, and visualization. Here we are investigating biological sources of cellular variation to identify the basis functions of a dimensionally-reduced, interpretable parameter space that represents integrated intracellular organization. We used the Allen Cell Structure Segmenter to create accurate 3D segmentations of cells and nuclei in a large, >100k single cell dataset and in multi- hour 3D timelapse movies. We fit extracted cell/nuclear shapes using spherical harmonic functions perform a PCA analysis. We analyzed the contributions of the first 5 primary axes of variation to describing this dimensionally-reduced cell and nuclear shape space. Each shape mode represented a different source of biological variation in hiPS cell colonies and occurred on a distinct timescale. The first two shape modes represented cell growth during the cell cycle and cell colony packing density occurring over several days. The next three axes of variation represent distinct aspects of cell/nuclear shape, such as how elongated these are in the XY plane, which occur over minute timescales due to constant interactions between neighboring cells. We are now applying these analyses to develop biophysical models of cell/nuclear shape and colony dynamics. This general analysis framework will be extended to each of the key intracellular structures in an integrative fashion.

Deregulation of Receptor Tyrosine Kinase (RTK) signaling underlies a large fraction of human cancers. These genetic perturbations lead to a variety of cell autonomous and non-cell autonomous effects ranging from cell proliferation to epithelial cell extrusion (e.g. Epithelial Defense Against Cancer, EDAC). Previous studies have shown that the temporal patterns of RTK-MAPK signaling (i.e. dynamics) can differentially regulate cell physiology. However, the role of signaling dynamics in mediating the effects of cancer driving mutations has not been systematically explored. Using live-cell imaging of signaling biosensors upon induction of oncogenic perturbations, we demonstrate that MAPK activity dynamics are decoded by the ADAM17-EGFR paracrine signaling axis to influence neighboring cell behaviors in epithelial monolayers. Specifically, sustained––but not pulsatile––ERK activity triggers amphiregulin (AREG) release and EGFR-ERK signaling in neighboring cells that coordinates migration and proliferation of neighboring cells to promote oncogenic cell extrusion. Interestingly, both oncogenic and neighboring cells require ERK signaling but with qualitatively different dynamics. Overall, we show that the temporal patterns of MAPK activity differentially regulate cell autonomous and non-cell autonomous effects of oncogene expression.

This session introduces basic concepts of imaging-based high-throughput screening (HTS) and high-throughput profiling assay development. Imaging-based HTS assays are designed to evaluate a discrete cellular process and produce a single, or small number of quantitative outputs. In contrast, imaging-based HTP assays measure dozens to thousands of features and provide highly multiplexed quantitative outputs.  Either type of approach may be used to evaluate the effects of chemicals or other perturbagens on cellular biology. Topics for this session include (but are not limited to) considerations for model selection, endpoint selection, imaging assay design, identification and use of positive control and reference treatments, methods for evaluating assay dynamic range and approaches for evaluating assay reproducibility. Examples of HCS assays for a variety of biological processes including nuclear receptor activation, oxidative stress, apoptosis, neurite outgrowth and others will be explored. In addition, this presentation will explore the basic concepts of machine learning classification in the context of HTP assays such as Cell Painting. Attendees will gain a basic foundational knowledge of guiding principles underlying the development of imaging-based HTS and HTP assays.  The views expressed in this presentation are those of the author and do not necessarily reflect USEPA policy.

The pharmaceuticals and biotechnology industries have made substantial progress in reducing the number of drugs that fail in clinical trials due to safety. This has been due, in part, to an improvement in the ability to identify problematic compounds at an earlier stage in the drug discovery process, using in vitro methods. One technology that is increasingly being used in this effort is high content analysis, (HCA).

This educational session will serve as an introduction to the field of HCA for toxicology. we will review how HCA can increase the efficiency of tox screening through the multiplexing of assays. There will be a special focus on how HCA, when combined with various forms of Artificial Intelligence, can be used for predictive toxicology. We will discuss how the technologies have been applied in various disease fields and the challenges associated with the implementation of these methods.

Learning Objectives: 

• Get an overview of the application of HCA for in vitro toxicology 
• Understand how the combination of HCA artifical intelligence can be beneficial for predictive toxicology 
• Gain insights into the challenges associated with the implementation of these methods

Fluorescence Resonance Energy Transfer (FRET) between fluorescence proteins has been implemented for a number of biosensors in which the donor and acceptor are linked in a single sensor.  For example, many sensors have been published to measure caspase activity in live cells using a single molecule consisting of a CFP donor linked via a caspase site to an YFP acceptor. In single molecule sensors FRET can be measured using either stimulated emission or Fluorescence Lifetime Imaging Microscopy (FLIM) because the ratio of donor to acceptor is one. However, to measure protein:protein interactions in live cells is more complicated because the relative concentration of donor and acceptor are unknown. Because fluorescence lifetime is independent of concentration, it is possible to use FLIM FRET to quantify binding between any two proteins in live cells. By successfully automating FLIM FRET assays for high throughput we enabled examining the combinatorial interactions between 4 anti-apoptosis proteins with 6 different pro-apoptotic binding partners and their modulation by 15 drugs at 5 concentrations each.

In this webinar we will discuss the benefits and limitations of different methods for measuring FRET in cells. I will describe the exceptional utility of this kind of data in early stage drug development. Also discussed will be the steps required to surmount the challenges posed when generating 1800 binding curves in a single experiment with live cells and how to interpret the data to provide the most useful single value to guide medicinal chemistry efforts. Finally, we will explore some of the unexpected insights garnered from our data examining inhibitors on anti-apoptosis proteins in live cells.  

Learning Objectives:

1. Understanding what FRET is and how it can be used to measure protein:protein interactions

2. Understanding how FRET can be measured in live cells using automated microscopy

As High Content Analysis (HCA) has matured and gained wider adoption as a quantitative research tool, the application space has grown and is no longer limited to a finite list of well defined assays performed in standard biological models. To account for this added complexity, a large focus has been placed on improving the flexibility and performance of analysis methods. Machine Learning is becoming ubiquitous and there are many examples these methods outperforming traditional methods for applications across many industries, scientific disciplines, and every day life. One would assume that technologies such as Deep Learning could provide similar benefits in HCA. But with any new technology, one has to weigh the potential benefits against the associated costs of adoption. We will present background information describing some of the fundamental differences between image analysis methods, considerations one should keep in mind when evaluating analysis methods, and some examples suggesting the improvements one could realize by employing technologies such as Deep Learning.

Learning Objectives

• Limitations of image processing methods and the advantages of machine learning for object segmentation
• Survey of machine learning approaches and key differences with respect to their application in HCA
• The impact of improved segmentation accuracy on the overall endpoint of High Content screens