The High-Content Analysis conference, taking place November 6-7 in Cambridge, MA, will address the advantages of phenotypic screening vs. target-based screening, and focuses on image and data analysis solutions, computational strategies, and assay development and subsequent target identification and validation for high content analysis and phenotypic screening.

Final Agenda

Stay on and attend (Tues-Wed): Phenotypic Screening  

Monday, November 6

7:15 am Conference Registration and Morning Coffee

Data & Image Analysis Techniques for High-Content & Phenotypic Screening

8:15 Chairperson’s Opening Remarks

Gustavo Rohde, Ph.D., Associate Professor, Department of Biomedical Engineering, Department of Electrical and Computer Engineering, University of Virginia


8:25 FEATURED PRESENTATION: Revolutionizing Diagnostics and Drug Discovery with Deep Learning

Philip Nelson, Ph.D. Director, Software Engineering, Google

Mike Ando, Ph.D., Research Scientist, Google

Recent advances in machine learning have opened up new avenues of exploration and new possibilities for making sense of biological data. In this talk, Philip details some of what’s new in machine learning, how those insights have been applied in areas as diverse as tagging photos, translating language, playing Go, and how Google is making this technology available at tensorflow.org and cloud.google.com/ml . He then describes some of the ways Google Research is applying machine learning in bio contexts including medical imaging, interpreting microscopy images, computational chemistry, genomics, and connectomics. This talk will also detail an innovative approach to deriving new biological hypotheses from phenotypic screens using embeddings from models pre-trained on hundreds of millions of consumer images.

9:25 Analyzing Phenotypic Assays Using Large Scale Data Integration

Stephen Reiling, Ph.D., Senior Scientist, Chemical Biology and Therapeutics, Novartis Institutes for BioMedical Research, Inc.

The talk will describe the creation and use of a large knowledge base assembled from in-house and published data relevant to the analysis of phenotypic screens. The knowledge base comprises small molecule activity data, systems biology knowledge, and relationships from text mining/natural language processing of article abstracts. The knowledge base can be accessed as a relational database but also as a large graph, which enables the use of novel analysis methods. Some examples of how this system can be used for the analysis of phenotypic screens will be highlighted.

9:55 Coffee Break with Exhibit and Poster Viewing

10:30 Generative Predictive Modeling for High Throughput Screening: Getting Interpretable Results from High Content Screens

Gustavo Rohde, Ph.D., Associate Professor, Department of Biomedical Engineering, Department of Electrical and Computer Engineering, University of Virginia

Image-based phenotypic screens typically rely on parametric numerical features to describe the state of each cellular measurement. In recent years, numerous modeling frameworks for image-based cell phenotypes. We will describe how non-parametric methods based on the continuity equation can be used to build non-parametric end to end systems that utilize all pixel information directly in statistical modeling. We will show that the modeling framework can not only obtain predictive models of higher accuracies, but due to the invertibility of the modeling framework, allow for biologically interpretable results.

11:00 Life beyond the Pixels: Phenotyping Using Machine Learning and Image Analysis Methods

Peter Horvath, Ph.D., Group Leader, Synthetic and Systems Biology Unit, Hungarian Academy of Sciences, BRC, Szeged; Finnish Distinguished Professor (FiDiPro) Fellow, Institute for Molecular Medicine Finland (FIMM)

I will give an overview of the computational steps in the analysis of single cell-based large-scale microscopy experiments. First, I will present a novel microscopic image correction method designed to eliminate vignetting and uneven background effects. Novel single-cell image segmentation methods will be presented using energy minimization methods. I will discuss the Advanced Cell Classifier (ACC) (www.cellclassifier.org), a machine learning software tool capable of identifying cellular phenotypes based on features extracted from the image. It provides an interface for a user to efficiently train machine learning methods to predict various phenotypes. Our recently developed single-cell isolation methods, based on laser-microcapturing and patch clamping, utilize the selection and extraction of specific cell(s) using the above machine learning models.

11:30 Luncheon Presentation (Sponsorship Opportunity Available) or Enjoy Lunch on Your Own

12:15 pm Session Break

1:30 Chairperson’s Remarks

Vance Lemmon, Ph.D., Walter G. Ross Distinguished Chair in Developmental Neuroscience, Professor of Neurological Surgery, The Miami Project to Cure Paralysis; Program Director in Computational Biology, Center for Computational Science, University of Miami Miller School of Medicine

1:35 Learning Potential Causal Relationships from High Content Images

Robert F. Murphy, Ph.D., Ray and Stephanie Lane Professor of Computational Biology and Professor of Biological Sciences, Biomedical Engineering, and Machine Learning; Head, Computational Biology Department, School of Computer Science, Carnegie Mellon University

We have previously described image analysis methods that can be used to “register” the patterns of different fluorescently-tagged proteins into a common framework such that they can be compared even if they were not acquired in the same cell. We have now developed approaches to identify potential causal relationships between different protein locations, for example, learning that a change in location of one protein in one cell region is always followed by a change in location of a different protein at a different time.

2:05 Accelerating the Therapeutic Discovery Pipeline with Image Analysis Solutions

Peter Antinozzi, Ph.D., Assistant Professor, Department of Biochemistry, Department of Internal Medicine (Section of Molecular Medicine), Department of Genomics and Personalized Medicine Research, Center of Diabetes Research, Center on Diabetes, Obesity, and Metabolism, Wake Forest University School of Medicine

Human clinical trials of skin injury therapeutics often rely on in-office exams to assess efficacy outcomes. Such personnel-intensive assessments typically require multiple trained medical professionals present at the exam to evaluate the skin healing with a standardized scoring system. Here we present a probabilistic classifier-based computational image analysis strategy to assess skin injury therapeutic efficacy. This image analysis approach classifies and maps healing across the wound and has the ability to spatially integrate with computationally-assessed histology measures from biopsy samples. The resulting computational assessments match outcomes determined from clinical exams and have the additional benefit of expanding the number of features to characterize therapeutic efficacy.

2:35 Drug Discovery via High-Throughput Imaging and Deep Learning

Berton Earnshaw, Ph.D., Director, Data Science Research, Recursion Pharmaceuticals

Recursion Pharmaceuticals high throughput screening pipeline produces many hundreds of thousands of multi-channel images of cells under a huge variety of genetic and chemical perturbations every week, at relatively little cost per image. By training deep convolutional neural networks on these highly-informative image sets, along with auxiliary expression and cheminformatic data, we algorithmically extract relevant features for discriminating phenotypes associated with disease and drug rescue. In my talk, I will describe how at Recursion we have used these techniques to discover over 30 potential therapeutic candidates.

3:05 Refreshment Break with Exhibit and Poster Viewing

3:45 High-Throughput Morphological Analysis of Complex Systems Using CellProfiler and CellProfiler Analyst

Beth Cimini, Ph.D., Computational Biologist, Broad Institute

Microscopy images contain rich and often untapped information about the state of cells, tissues, and organisms. Using the open-source software tools we produce for high throughput image analysis (CellProfiler) and machine learning (CellProfiler Analyst), we have sought to increase both the complexity of systems we can analyze and the sophistication of the resulting morphological profiles. Recent work has allowed us to push this work into the analysis of 3D structures as well as cell types lacking a "typical" fibroblast morphology. Ultimately, we aim to make microscopy images as computable as other sources of genomic and chemical information.

Identifying & Validating Drug Targets from Phenotypic Screening

4:15 FEATURED PRESENTATION: Hit Triage and Mechanism Validation for Phenotypic Screening: Considerations and Strategies

Fabien Vincent, Ph.D., Associate Research Fellow, Assay Development and Pharmacology, Hit Discovery and Lead Profiling, Pfizer Global Research & Development

Phenotypic drug discovery approaches can positively affect the translation of preclinical findings to patients. However, significant differences exist between target-based and phenotypic drug discovery, prompting a need to re-assess and rethink our strategies and processes to most effectively prosecute phenotypic projects. Accordingly, the hit triage and validation process was re-evaluated in light of the unique characteristics of phenotypic screening. Key considerations and specific strategies will be shared and exemplified by in house and literature case studies.

4:45 Using Machine Learning, Phenotypic Assays and Biochemical Profiling to Identify Drug Targets and Anti-Targets

Vance Lemmon, Ph.D., Walter G. Ross Distinguished Chair in Developmental Neuroscience, Professor of Neurological Surgery, The Miami Project to Cure Paralysis; Program Director in Computational Biology, Center for Computational Science, University of Miami Miller School of Medicine

While phenotypic assays on live cells are very efficient at finding probes that produce desired outcomes it is often difficult or impossible to identify the probes’ targets. Conventional high throughput binding assays are extremely effective at identifying compounds that bind to specific targets but these assays do not ensure the compounds work well in live cells and it can take years to identify off-target interactions that may disrupt efficacy or cause undesirable side effects. By using highly annotated libraries of compounds that have been screened on hundreds of enzymes, for example kinases, in phenotypic assays it is possible to identify targets and anti-targets (that prevent the desired phenotypic outcome) using machine learning approaches. This drug discovery pipeline is extremely efficient and uncovers synergies between targets that can dramatically increase efficacy.

5:15 Welcome Reception with Exhibit and Poster Viewing

6:15 Close of Day

Tuesday, November 7

 

8:00 am Breakfast Breakout Roundtable Discussions

Concurrent breakout discussion groups are interactive, guided discussions hosted by a facilitator to discuss some of the key issues presented earlier in the day’s sessions. Delegates will join a table of interest and become an active part of the discussion at hand. To get the most out of this interactive session and format please come prepared to share examples from your work, vet some ideas with your peers, be a part of group interrogation and problem solving, and, most importantly, participate in active idea sharing.

Table 1: Deep Learning and High-Throughput Imaging

Moderator: Berton Earnshaw, Ph.D., Director, Data Science Research, Recursion Pharmaceuticals

Table 2: Using Microphysiological Systems in Preclinical Drug Development

Moderator: Michael L. Shuler, Ph.D., Samuel B. Eckert Professor, and J. & M. McCormick Chair, Chemical & Biomedical Engineering, Cornell University; Director of Cornell’s Nanobiotechnology Center

  • How authentic do the systems need to be to be useful?
  • Could an all human microphysiological system replace animals in preclinical studies in the next 10 years?
  • Could microphysiological models be used to predict drug-drug interactions?

Table 3: Artificial Intelligence and the Pathologist

Moderator: Peter Antinozzi, Ph.D., Assistant Professor, Department of Biochemistry, Department of Internal Medicine (Section of Molecular Medicine), Department of Genomics and Personalized Medicine Research, Center of Diabetes Research, Center on Diabetes, Obesity, and Metabolism, Wake Forest University School of Medicine

  • What is the likelihood of raw computing power and AI algorithms to emulate (replace) a pathologist's interpretation of diagnostic images?
  • What are the primary technical challenges?
  • What are the regulatory and societal barriers?

Table 4: Bioprinting 3D Structures

Moderator: Min Jae Song, Research Fellow, National Eye Institute, National Institute of Health

Table 5: Comparison of Cell Phenotypes across Different Cell Types or Instrumentation

Moderator: Robert F. Murphy, Ph.D., Ray and Stephanie Lane Professor of Computational Biology and Professor of Biological Sciences, Biomedical Engineering, and Machine Learning; Head, Computational Biology Department, School of Computer Science, Carnegie Mellon University

  • Have you encountered a need to compare the specific phenotypes observed in one assay or screen with phenotypes from another screen performed using different equipment?
  • What approaches would work for this task?
  • How about comparing phenotypes across different cell types, either when images were taken with the same or different instruments?
 

The Future of High-Content Analysis and Phenotypic Screening

8:55 Chairperson’s Remarks

Robert F. Murphy, Ph.D., Ray and Stephanie Lane Professor of Computational Biology and Professor of Biological Sciences, Biomedical Engineering, and Machine Learning; Head, Computational Biology Department, School of Computer Science, Carnegie Mellon University

9:00 Breakout Roundtable Report-Outs

9:30 Key Learnings for Success from 5 Years of Phenotypic Drug Discovery at Novartis

Christophe Antczak, Ph.D., Senior Investigator, Hit Discovery Sciences, Chemical Biology & Therapeutics, Novartis Institutes for BioMedical Research, Novartis Institutes for BioMedical Research

Phenotypic drug discovery (PDD) has undergone a renaissance in recent years, due to its potential to deliver first-in-class drugs to treat diseases lacking validated targets or clearly understood molecular basis. Teams at Novartis have been successful in navigating the challenges associated with PDD, and we sought to identify any trend in their approach that increases the likelihood of success of a PDD project. A retrospective analysis of 5 years of PDD at Novartis provides guidelines expected to improve the success rate of teams involved in PDD. We provide an overview of the results of this analysis, highlighting key learnings.

10:00 Coffee Break with Exhibit and Poster Viewing

10:40 PANEL DISCUSSION: High-Content Analysis and Phenotypic Screening Trends and Bottlenecks

Moderator: Will Marshall, Product Manager, GE Healthcare

Jonathan A. Lee, Ph.D., Senior Research Advisor, Quantitative Biology, Eli Lilly and Company

Christophe Antczak, Ph.D., Senior Investigator, Hit Discovery Sciences, Chemical Biology & Therapeutics, Novartis Institutes for BioMedical Research, Novartis Institutes for BioMedical Research

Berton Earnshaw, Ph.D., Director, Data Science Research, Recursion Pharmaceuticals

This panel will discuss current trends and bottlenecks in phenotypic screening: What are the next challenges and opportunities that lie ahead for phenotypic screening? What are the current limitations to existing technologies and where will the next breakthroughs take place? What are current challenges in assay development?

11:40 Close of Conference

Stay on and attend (Tues-Wed): Phenotypic Screening