Bioinformatics Training & Education Program

AI Models of Cancer and Precision Medicine: Building a Mind for Cancer

The long-term objective of the Ideker Lab is to create artificially intelligent, mechanistic models of cancer and neurodegenerative diseases for translation of patient data to precision diagnosis and treatment. We seek to advance this goal by addressing fundamental questions in the field: What are the genetic and molecular networks that promote disease, and how do we best chart these? How do we use knowledge of these networks in intelligent systems for predicting the effects of genotype on phenotype? – Ideker Lab, https://idekerlab.ucsd.edu/research/cancer/

This webinar will be recorded and made available on the BTEP web site: https://bioinformatics.ccr.cancer.gov/btep/btep-video-archive-of-past-classes/ within 48 hours after the event ends. 

Workshop Description: The application of AI to cancer research holds promise to accelerate new discoveries, enable early detection, improve diagnosis, and spur development of new therapies for cancer. Machine learning and other forms of AI have made a significant impact in some areas of cancer research, but the full promise of data-driven approaches has been elusive. While there are important ongoing efforts to collect and produce large, well-annotated datasets to support the training of robust deep learning models, the heterogeneity and complexity of cancer, along with privacy and bias concerns, continues to limit the application of AI methods to many critical areas of cancer research. There is a need for foundational advances in machine learning that can operate on incomplete, noisy, unbalanced and/or biased data across the cancer research continuum.

The goals of this workshop are to (1) examine the state of the science for AI methods designed to operate on noisy, complex, or low-dimensional data, (2) explore how these methods may be applied to key areas of cancer research, and (3) discuss processes for identifying the biological questions that will motivate further advances in machine learning. This workshop will highlight the importance of leveraging advances across fields to accelerate cancer research and discovery through AI.

Workshop Chairs:

Caroline Uhler, Ph.D. (MIT and Broad Institute)

Olivier Gevaert, Ph.D. (Stanford University)

NCI Planning Committee:

Juli Klemm, Ph.D.

Jennifer Couch, Ph.D.

Sean Hanlon, Ph.D.

Natalie Abrams, Ph.D.

Keyvan Farahani, Ph.D.

Emily Greenspan, Ph.D.

Paul Han, M.D., M.A., M.P.H.

Roxanne Jensen, Ph.D.

Jerry Li, M.D., Ph.D.

A summary of the planned workshop sessions and participants is provided below. A detailed agenda with speakers and presentation titles will be posted ahead of the meeting.

Welcome and Opening Comments

• National Cancer Institute
• Caroline Uhler, MIT and Broad Institute

Session 1: Integrating classical structure prediction with machine learning towards drug discovery

Session Chair: Trey Ideker, UCSD

This session will focus on expanding the field of structure prediction to incorporate multiple data modalities and layers of biological structure beyond the protein, as well as meta-learning for identifying targets for drug discovery.

Speakers:

• Anima Anandkumar, Cal Tech and NVIDIA
• Andrej Sali, UCSF
• Jure Leskovec, Stanford

Panelists:

• Rick Stevens, Argonne National Laboratory
• Sergey Ovchinnikov, Harvard

Session 2: Chemical, genetic, and mechanical perturbations for understanding mechanisms in cancer: Extrapolating beyond existing data

Session Chair: Fabian Theis, Helmholtz Munich

In this session, researchers will discuss the use of large-scale perturbation data for causal modeling, combining representation learning with perturbation approaches, and methods to extrapolate beyond existing perturbation data.

Speakers:

• Yoshua Bengio, Université de Montréal
• GV Shivashankar, ETH Zurich
• Smita Krishnaswamy, Yale

Panelists:

• Paquita Vazquez, Broad Institute
• Byung-Jun Yoon, Texas A&M University and Brookhaven National Laboratory

   

Session 3: Multimodal learning in data limited contexts: Leveraging tissue-level data for understanding cell-cell interactions in cancer

Session chair: Dana Pe’er, Memorial Sloan Kettering

This session will focus on multimodal learning in data limited contexts, including cell-cell interactions and predicting outcomes. Dealing with imbalances across multimodal data sets and foundational models will also be discussed.

Speakers:

• Elena Fertig, Johns Hopkins
• Elham Azizi, Columbia
• Livnat Jerby, Stanford

Panelists:

• Marianna Rapsomaniki, IBM Research
• Arjun Krishnan, University of Colorado

   

Session 4: Making use of large-scale, structured clinical research data and image repositories

Session chair: Ziad Obermeyer, UC Berkeley

In this session, researchers will discuss the use of large-scale clinical research data for machine learning models. Discussion topics include the use of synthetic data, considerations of bias, generalizable models, and development of digital twins.

Speakers:

• Chris Probert, InSitro
• James Zou, Stanford
• Mihaela van der Schaar, University of Cambridge

Panelists:

• Lily Peng, Verily
• Matthew Lungren, Microsoft/UCSF

Session 5: Improving modeling of real-world evidence data in clinical research and clinical trial design

Session chair: Tianxi Cai, Harvard

This session will focus on real-world evidence (RWE) data modeling, including issues associated with RWE data such as electronic health record coding and unbalanced data, towards the development of clinical trials.

Speakers:

• Sean Khozin, MIT
• Limor Appelbaum, Beth Israel Deaconess
• Ryan Copping, Genentech

Panelists:

• Donna Rivera, FDA
• Khaled El Emam, University of Ottawa

   

Session 6: Cross-cutting discussion with session chairs

Session chair: Olivier Gevaert, Stanford University

Discussion of the approaches and challenges identified during the workshop and opportunities for the future.

Panelists:

• Caroline Uhler, MIT and Broad Institute
• Trey Ideker, UCSD
• Dana Pe’er, Memorial Sloan Kettering
• Ziad Obermeyer, UC Berkeley
• Tianxi Cai, Harvard

This class provides a basic overview of the methods used to visualize the association among two or more quantitative variables. This class will focus on scatterplots, scatterplot matrix, and visualizing paired data. Participants are expected to have taken the Introduction to Data Visualization in R: ggplot class. Participants are encouraged to install R, RStudio, and the tidyverse package, before the webinar so that they can follow along with the instructor.

Python is a programming language used for data science, specifically: data analysis, statistical analysis, and visualization of results. This class will demonstrate integrated development and learning (IDE) platforms for learning Python, the fundamentals of Python coding, and why it is advantageous to develop these skills.  The session will feature the following IDEs: Google Colaboratory: Jupyter Notebook; and Anaconda’s: Spyder, Jupyter Notebook, and JupyterLab.  It will also provide an overview of programming constructs needed to learn Python. Finally, this class will demonstrate why these skills can boost productivity, rigor, and transparency in reporting research findings

Azimuth is a web application that uses an annotated reference dataset to automate the processing, analysis, and interpretation of a new single-cell RNA-seq experiment. Azimuth leverages a 'reference-based mapping' pipeline that inputs a counts matrix of gene expression in single cells, and performs normalization, visualization, cell annotation, and differential expression (biomarker discovery). All results can be explored within the app, and easily downloaded for additional downstream analysis. - Satija Lab

The development of Azimuth is led by the New York Genome Center Mapping Component as part of the NIH Human Biomolecular Atlas Project (HuBMAP).

This webinar will be recorded and made available on the BTEP web site: https://bioinformatics.ccr.cancer.gov/btep/btep-video-archive-of-past-classes/ within 48 hours after the event ends. 

Wondering why you should use R for data visualization? Lesson 1 of the Data Visualization with R course series will address this question and introduce the various plot types that will be generated throughout the course. Lesson 1 will also showcase related plots that you will be able to create in the future using the foundational skills gained over the next 5 lessons.

This will not be a hands-on lesson so no coding just yet. The hands-on portion of this series will start with lesson 2, Getting Started with ggplot2. 

This lesson is the first lesson of a multi-lesson course series. Registering here will register you for the entire course series.

IMPORTANT: You do not need to download or install any software to participate in the course. This course will be taught on the DNAnexus platform. Every learner will need to create a free DNAnexus account at https://dnanexus.com. After you have created your DNAnexus account, please complete this form. If you fail to complete the form, we will not be able to give you access to the course on DNAnexus.

In this hands-on virtual lab, the participants will familiarize themselves with Deep Learning concepts and techniques, using MATLAB Online to train deep neural networks on GPUs in the cloud, create deep learning models from scratch for images and signal data, explore pretrained models and use transfer learning, import and export models from Python frameworks such as Keras and PyTorch, as well as automatically generate code for embedded targets. Deep Learning can achieve state-of-the-art accuracy when it comes to complex problems such as image classification or developing predictive models for signal processing applications. Deep Learning outperforms humans in some tasks like classifying objects in images. Cancer researchers are using deep learning to automatically detect cancer cells, for example.

Lesson 2 of the Data Visualization with R course series will focus on the basics of ggplot2, including the grammar of graphics philosophy and its application. This lesson will provide a hands on introduction to the ggplot2 syntax, geom functions, mapping and aesthetics, and plot layering.

Registering for lesson 1 of this course series will enroll you in the entire course series.

This seminar will introduce the NIH Comparative Genomics Resource (CGR), an NIH-funded, multi-year NLM project to establish an ecosystem to facilitate reliable comparative genomics analyses for all eukaryotic organisms in collaboration with the genomics community. The project’s vision is to maximize the biomedical impact of eukaryotic research organisms and their genomic data resources to meet emerging research needs for human health. To achieve this, NCBI is providing high-value data and assorted tools compatible with community-provided resources.