Breakthroughs require heart

Team Lead: Dennis Kennetz

Summary: This proposed project will be to write a server written in golang which will maintain files present versus needed via filesystem lookups and database queries based on active/open cases. If a file is needed but not present, the server will initiate a copy and update state once copy completes. If the file is present but not needed, it will archive the file by initiating a copy and subsequently a delete. It will also maintain the current state of the filesystem to monitor a capacity threshold so files will not be copied unless there is sufficient space. Finally, it will parallelize these processes via lightweight goroutines for the copy operations.

Skills Needed: Go, SQL, concurrent programming, data modeling.

Team Lead: Subodh Selukar

Summary: The Kaplan-Meier curve is a well-recognized tool for visualizing the survival function for time-to-event data. Multiple Kaplan-Meier curves can be plotted side-by-side to compare groups, but this approach can fail when subjects are able to change between groups after baseline. For example, Graft-versus-Host disease (GVHD) is an important complication following allogeneic hematopoietic cell transplantation (alloHCT), and assessing the association between occurrence of GVHD and survival outcomes is frequently of interest. One possible analysis could assign study subjects to groups of "ever GVHD" and "never GVHD" from baseline, but this assignment could bias the analysis because subjects experience GVHD after alloHCT, and the time between alloHCT and occurrence of GVHD is misattributed in the "ever GVHD" baseline group. An alternative analysis would allow for time-varying group membership, where study subjects could move from the "never GVHD" group to "ever GVHD" after first experiencing GVHD.

A recent development in this area proposed how to construct survival curves for user-specified group trajectories: e.g., the survival curve for a hypothetical group that experiences GVHD 30 days after alloHCT. In this project, we will implement this method in an R Shiny app to develop an interactive tool for visualizing survival curves with time-varying group membership. Users will be able to input data and desired trajectories and see the estimated curves of interest. We believe this tool may have important potential in the clinic to communicate risks and inform personalized treatment regimens.

Skills Needed: Experience with R, familiarity with survival analysis.

Team Lead: Ragha Srinivasan

Summary:

Background

SCCRIP (Sickle Cell Clinical Research and Intervention Program) established a longitudinal clinical cohort of Patients with Sickle Cell Disease (SCD) in 2014 in the St. Jude Clinical Hematology department.

Since SCCRIP began in 2014, researchers and physicians at institutions across the country have reached out to express interest in contributing data for SCCRIP. Currently SCCRIP affiliate or collaborating sites include four hospitals: Methodist/Le Bonheur Hospital (Memphis, TN); Children’s Hospital of Illinois (Peoria, IL); Novant Health (Charlotte NC); Our Lady of the Lake (Baton Rouge, LA). These hospitals contribute clinical data to SCCRIP by extracting data from their EMR and uploading the extracts to SCCRIP sFTP sites. SCCRIP data management staff review and audit the data extracts as they are made available and download all data annually to include in the data freeze. This approach to the transfer and harmonization requires significant resources to develop and maintain. Some sites who would like to participate in and contribute data to SCCRIP are unable to meet these requirements.

About OMOP CDM

The Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM) is an open community data standard, designed to standardize the structure and content of observational data and to enable efficient analyses that can produce reliable evidence.

Opportunity

SCCRIP currently has 1300 enrollees. One collaborator for SCCRIP has longitudinal data for 600 SCD patients in OMOP CDM format.

Transitioning SCCRIP data to an industry standard Common Data Model (CDM) is currently being evaluated at St. Jude, though the conversion may take until 2024/2025 to complete. Therefore, as an interim solution, to help accelerate research in Sickle Cell Disease, we are expanding the SCCRIP cohort through acquisition and conversion of collaborator's data in OMOP CDM format to SCCRIP.

Objectives

• Establish infrastructure to characterize data and perform data quality checks of data received in OMOP CDM format.
• Create an ETL (Extract, Transform, Load) process to convert affiliate OMOP CDM data to the SCCRIP Registry.
• Identify St. Jude Employees trained in OMOP CDM and build technical know-how among investigators and IT to generate evidence using the OHDSI platform in clinical settings.

Skills Needed: Networking, R, OHDSI tools, PostgreSQL/SQL, SAS, medical vocabulary such as LOINC, RxNORM, SNOMED, ICD-10.

Team Lead: Vyoma Sheth

Summary: Gene regulatory networks (GRNs) are powerful for understanding biological relationships. A significant volume of biomedical literature provides information on gene regulation at various levels. There are many NLP models out there that focus on biomedical literature mining and text generation like BioGPT/ Chili Bot. Still, these NLP models lack integration of details at molecular and phenotypic levels. Including this information in GRNs would provide a more holistic view of gene regulation at various levels and would be insightful for researchers to choose a target.

• Step 1: Use BioGPT and Chili Bot to extract the PubMed literature abstracts focusing on gene regulation using keyword search.
• Step 2: Build an AI model (NLP) to annotate all the genes that provide information on the regulation of other genes in simple statements derived from the output of Chili Bot/BioGPT, GO terms, and database resources (TTrust, etc).
• Step 3: Associate GRNs with phenotype information from clinically available GWAS data.
• Step 4: Deploy the model as a web interface/GUI.

Skills Needed: Biostatistics expertise, machine learning, web development, Python, GRN/biological expertise.

Team Lead: Hadi Hosseini

Summary: The ability to automatically detect, classify, calculate the size, number, and study membrane properties and cellular processes are critically important. Manually analyzing cells in microscopy images or with custom-made scripts make analysis time-consuming and results difficult to compare across studies. Most of this analysis is done by controlling the light intensity and filtering patterns based on Gaussian distribution. It causes the loss of some informative pixels in the microscopy images, with low resolution or blurry images.

Based on new development of artificial intelligence in computer vision in recent years, and the use of those algorithms in biomedical science, we have a plan to develop an automated tool based on deep learning techniques and the Mask R-CNN model to analyze microscopy images. The goals of this tool are to quantify and segment cells and common biological patterns inside the cells with high speed and precision. The shape of cells and other morphology information, like the localization of nuclei inside the cells, will be gathered from the segmented cells. The output of the designed model will be automatically detected, categorized, and count categorized cells to perform population analysis.

Ideally, this tool would have a GUI or reporting component to relay all information, filter by population or other morphological features, etc. The performance of our model will be tested using general metrics: general mean average precision (mAP) for detection, and precision, recall, F1-score, and accuracy for categorizing and counting.

Skills Needed: Machine/deep learning skills, biostatistics, experience in image processing with deep learning models, Python (Scikit-learn, TensorFlow, PyTorch, Matplotlib, Django, FastAPI), JavaScript (D3.js, web frameworks, database).

Team Lead: Charlie Wright

Summary: The High-Performance Computing Facility (HPCF) is an important resource for scientific computing at St. Jude. The HPCF provides infrastructure allowing advanced computation and handling of large datasets. The needs of this resource are campus-wide, but vary dramatically at the level of departments, labs, and individual users. Additionally, the usage of HPCF fluctuates by time of day and time of year.

Understanding the utilization of the HPCF is critical for proper allocation of resources, and for assessment of current infrastructure needs.

There is an unmet need to systematically explore, visualize, and interpret usage summary statistics. Usage reports can be generated but are currently not efficiently shared. A dashboard to display periodically updated data would be beneficial. Additionally, it may be useful to apply machine learning to forecast/predict the surge in demands of HPCF resources. By training a model on historical usage information, a prediction of demand would help mitigate strain on resources.

We propose the using the R framework to efficiently summarize, display, and apply machine learning models on the usage data. We would further create a dashboard in R/Shiny to display the results.

Skills Needed: Machine learning, programming (likely in R), dashboard development (likely in Shiny), familiarity with the St. Jude HPC and/or the LSF scheduler.

Team Lead: Jason Ochoada

Summary: Through the CCI team the department has been strategically acquiring small molecules with known biological activities against one or more gene targets. This collection of molecules is known as the CBT Mechanism of Action compound set or library. We would like to build upon this set which will continue to grow. These compounds will be screened against a number of target and cellular assays. These assays will provide a set of "active" compounds in that assay. Given the set of compounds which are active we would like to identify and map the gene targets and pathways which may be involved. Additionally, the ability to visualize the activity of each pathway to the user would be useful. In this case there is no existing solution. It would add new capability for the department.

Skills Needed: Pathway analysis, data viz in R or python, network analysis, drug/compound screening.

Team Lead: Felipe Segato Dezem

Summary: The aim of this project is to create an analysis pipeline to be able to perform a differential comparison of two or more (comparable) whole-slide, multiplexed spatial imaging samples (e.g PhenoCycler, CosMX, or any other image-based spatial assay). The proposed workflow includes QC, normalization, cell-type phenotyping based on gene/proteins used in the assay and a statistical model to compare different groups.

Skills Needed: Machine learning, R/Python, git, microscopy/spatial analysis.

Team Lead: Michael Root

Summary: This is a large, multi-faceted project that may be composed of sub-teams.

IDMS / NAPARI. This falls under Sub-Challenge 1: IDMS Front-end Dashboard Prototype

The goal here is to provide an interface within an existing open-source image analysis tool, napari, to read in data from our internal IDMS. This will showcase integration of the IDMS with a 3rd party imaging tool via the exposed REST API. The main work would be in napari and showcase how an interface would be constructed within Napari to browse data from the IDMS. There are 4 parts to the challenge with part 1 being the base requirement and parts 2, 3 and 4 added options if time allows during the hackathon. Accomplishing all 4 parts during the hackathon would be the ultimate goal and quite an achievement.

• Part 1. Display user, project, and image collection information within a custom napari interface.
• Part 2. Display image within napari from the backend cbi_renderer via the IDMS.
• Part 3. Display image within napari by specifying a 2D or 3D Region of Interest (ROI) thereby exercising additional API calls on the backend. If this is accomplished, then the napari/IDMS plugin would immediately be useful for many real applications.
• Part 4. Provide functionality to access a full image pyramid via napari-supported plugins and asynchronous data preparation. The data preparation step is triggered from within the new napari plugin (dashboard), routed from IDMS to a process that generates the pyramid and knows that it is talking to the Renderer tool.

This falls under Sub-Challenge 1: IDMS Front-end Dashboard Prototype as an alternative solution to the one proposed above using napari.

The goal here is to provide a frontend prototype using React and Material UI to the IDMS backend. Dummy data is permitted to be created and used for construction and demonstration of envisioned dashboard. Project team members can utilize the IDMS and CBI_Renderer backend (REST API) to populate elements of the prototype dashboard. The dashboard will include the following visual elements and sub elements:

• a. Login / Logout Navigation
• b. Header / Footer sections / navigation
• c. Users list of projects and corresponding image collections
• d. Image collection detail meta data view
• e. Toolbar showing actions that can be taken on an image collection such as duplicate, transform, delete, edit meta data, etc.
• f. Image view window
• g. Image view toolbar showing actions that can be taken on the image
• h. Other elements the team can reasonably envision

Skills Needed: Python, REST APIs, interface design, image analysis, dask, zarr, napari expertise, and interest in big data.

Team Lead: Daniel Darnell

Summary: This projects seeks to establish a modern UI wrapper for at least two ML models found in the BioSeq-BLM package. The existing website for BioSeq-BLM is very outdated and not user-friendly. The main project goals will be: 1) Create interactivity within the notebook via Voila, 2) Explain/document parameters and affect(s) on the resultant model 3) graphically display results of the model built (which will update interactively as parameters are altered). Stretch goal(s) will be to incorporate more models from BioSeq-BLM.

Skills Needed: Machine learning, BioSeq-BLM, python, jupyter, Voila.

Team Lead: Jaimin Patel

Summary: AI-powered chatbots like ChatGPT can revolutionize how people perform day-to-day jobs, and cancer research is no exception. Researchers can find answers to complex questions within minutes without spending hours on PubMed to find any specific papers or go through several relevant research papers piling up on their desks. But private usage of public chatbots like ChatGPT for research can lead to potential risks of leaking confidential data to the internet.

We will demonstrate how to host a research-specific GPT or BERT model for internal usage at St. Jude to address this problem. For example, BioGPT is trained on biomedical data and PubMedBERT on PubMed data. So, this project will focus on comparing different models, hosting one, and providing a user-friendly UI like ChatGPT. We will also develop API capability, so other research labs in St. Jude can utilize the model and develop new applications.

Because these open-source models like GPT3 and BioGPT are trained only on public data, it's missing a large amount of valuable data for researchers, for example, unpublished work, restricted access papers, and proprietary data. This project will also explore if we can further train one of these pre-trained models on private or institutional data and make it available internally as JudeGPT. It will also open up an avenue to make models specific for different research interests, for example, KinaseGPT or LukemiaGPT. This project will also investigate the benefits and pitfalls of segregating various research topics in numerous models or using a singular model to power all research needs like ChatGPT.

Skills Needed: Machine learning with NLP, Python familiarity with scikit-learn, Pytorch, frontend development with JS frameworks like React, backend development in python with framework like Django or FastAPI.

Team Lead: Liam Hallada

Summary: Computer vision is a rapidly changing field that is becoming an essential component of biological research. As exemplified by current discourse surrounding Artificial Intelligence with image generation, it is difficult to keep up with this evolving technology. Computer scientists across the world have been producing a sprawling set of tools to implement at each stage of image analysis. Because of this sprawl, we require an image analysis platform capable of integrating the latest tools in computer vision. This flexibility is usually achieved at the cost of user-friendly interfaces.

Napari is an open-source program that addresses this complicated issue with a python interface. Napari is a particularly capable at integrating image analysis as a single environment that can handle machine learning enhanced preprocessing, segmentation, and analysis of datasets larger than 40 GB. Because Napari is a relatively new program, it does not have the support and breadth of resources compared to other programs.

We will work with St. Jude High Performance Computing to set up a user-friendly Napari environment that is capable of machine learning augmented image analysis. To make Napari as user friendly as possible we will assemble the base code for how to import data from our servers, visualize complex imaging datasets, export results back to the server, and basic batch processing of data folders. We will set up this Napari environment with machine learning-enhanced image segmentation by installing plugins for Ilastik and CellPose that have already been generated. To further expand the toolbox of advanced imaging tools available to users, we will set up a plugin to integrate the Advanced Normalization Tools (ANTs) program that is widely used for image registration, segmentation, and analysis. We will need a team of people to code in python to achieve this goal, and we will need to test the user interface on researchers with a range of computer program experience. Altogether we will provide a final distribution of Napari that is ready for St. Jude users to implement machine-learning augmented image analysis.

Skills Needed: Python, image segmentation, experience with napari/CellPose/Ilastik/ANTs, and plug-in development.

Team Lead: Susanna Downing

Summary: This project aims to develop high-quality, flexible R Shiny modules for common plots used in bioinformatics analyses (e.g. volcano plots, scatter plots, box/violin plots, heatmaps, line plots, etc). These modules will provide interactive plots (with (gg)plotly or d3.js) with full customization over the aesthetics via a compact series of inputs that users can tweak on the fly.

This project will dramatically decrease the time and thought it takes to develop a Shiny application that utilizes common bioinformatics data structures and analyses by providing drop-in ready modules that require only a few lines of code and provide all the functionality a developer could ever want. These composable modules will be particularly useful in the context of Rmd notebooks hosted on with the Shiny runtime, as they’ll allow interactive exploration of analysis results and data structures with appropriate contextual information provided by the analyst.

Skills Needed: R, Shiny, ggplot2, plotly, common R bioinformatics data structures, and general data viz experience.

Team Lead: Jared Andrews

Summary: This project is centered around multi-modal data integration to identify putative core regulatory circuitry (CRCs) on a sample-by-sample basis. In short, given a BED file of ATAC peaks (or H3K27ac peaks), a BED file of SE calls from ROSE, and an expression matrix of counts/TPMs for a given sample, identify putative CRCs based on bi-directional network analysis.

Current tools to perform these steps (coltron, CRCmapper, CRC) are no longer maintained, difficult to install/run, and/or out of date (python 2, old annotations/motif databases, etc). This project will create a new, simplified python package that incorporates these methods to be used on semi-processed datasets - matched ATAC-seq (optional), RNA-seq (optional), and H3K27ac SE calls for a given sample.

In addition, it'd be useful to building reports/dashboards (dash, Shiny, etc) to effectively visualize the resulting CRCs and their downstream gene targets, potentially with GO/pathway enrichment incorporated. Lastly, comparing CRCs between samples/groups in some way would help reveal variable CRC members, true "core" members, and relate CRC membership to specific phenotypic states.

A solution to this problem would allow us to include putative CRC identification into our standard workflow for sample analysis. It would also kickstart the field with a much-needed update to these methodologies and provide a mechanism to better integrate these multi-modal datasets via cohesive, holistic analyses.

Skills Needed: Python, dash, network analysis, interval operations in python (pyranges, etc), motif analysis, python package development, and basic epigenomics knowledge.

Team Lead: Lei Li

Summary: Challenges: Single-cell datasets are usually analyzed by bioinformaticians, and they report results/figures to biologists, who is the key person that drives the project. However, time for both parties has been wasted during communications in some routine tasks, for example, DE genes between two subsets, gene expression of a few genes, and so on. Current solutions for single-cell analysis and visualization, especially linking BCR/TCR with muti-omics, either python-based (jupyter notebook using python packages) or R-based (RStudio using R packages), require extensive computer background, which is very difficult for biologists. For example, scRepertoire, a R package can link V(D)J with multi-omics, requires extensive computer background. Existing GUI tools more focus on scRNA-seq analysis and rarely compatible with V(D)J sequences. For example, iSEE, a R/shiny interface for single-cell visualization, KANA, a Bioconductor/R-centric analysis approach, and cellxgene, a python-based toolkit.

Plan: We will build an extension on top of existing software developed for V(D)J analysis (VGenes, https://wilsonimmunologylab.github.io/VGenes/). The extension will add support for scRNA-seq results and link them with V(D)J records. This software will be: 1) Python-based, because there are lots of python packages for single-cell analysis (e.g. SCANPY), and Python-based GUI development is more powerful (e.g. PyQt). The software is a standalone software and can be easily installed. We will also add support for R-processed results later. 2) To start, we will focus on data visualization. Bioinformaticians will process the raw data (QC, pre-process, integration, annotation), send the processed data object file to biologists, then biologists can explore their data using this GUI software at will on their own computer. For example, they can check the DE genes between any two cell populations by themselves or investigate the expression of any gene with a few clicks. 3) Future work will could add additional analysis capabilities to allow biologists to explore more details about their data.

Skills Needed: Python, PyQT5, Scanpy, Anndata, familiarity with scRNA-seq analysis.