St. Jude KIDS25 BioHackathon Challenges

Title: 
Automated Tumor Volume Estimation from MRI Data

Submitter: 
Stephen Mack (Developmental Neurobiology)

Summary: 
Over the past year, St Jude has generated thousands of MRI images of mice bearing various tumors including brain cancers. Unfortunately, tumor volumes are all quantified by hand. This impacts not only speed of data return, but potentially introduces unwanted biases in the analyses.

Benefit: 
Development of a "AI" tool to quantify tumor volumes would accelerate return of data to experimenters and reduce human time quantifying tumors manually. This would be beneficial to the brain tumor community, and could be extended into a future package that targets solid tumors. 

Tools: N/A

Test Data: 
We could give you hundreds-thousands of MRI images of animals with tumors to develop a training and prediction dataset.

Do any of the proposed datasets contain PHI, sensitive information, or otherwise confidential data?  No

Title: 
Proteomic Enrichment Analysis Interface

Submitter: 
Fabio Demontis (Developmental Neurobiology)

Summary: 
It would be great to have a user-friendly interface to compare a list of proteins of interest against the AlphaFold database (v2 or v3) with the purpose of identifying structural features, motifs, and domains (or combinations thereof) that are over/under-represented compared to the entire proteome or a reference list of proteins. 

Benefit: 
It would benefit all groups that wish to identify structural features enriched in a subset of proteins of interest.

Recommended software: N/A

Test Data: N/A

Do any of the proposed datasets contain PHI, sensitive information, or otherwise confidential data?  No

Title: 
Omics Dataset Validation Prior to Download

Submitter: 
Maycon Marção (Developmental Neurobiology)

Summary: 
I want to make it much easier to evaluate publicly available omics datasets—like single-cell RNA-seq, spatial transcriptomics, and DNA methylation arrays—before we actually download anything. The main source for these datasets will be the Gene Expression Omnibus (GEO). The goal is to avoid spending time and resources on downloading large amounts of data, only to find out that most of it isn’t useful for our research. For example, we want to quickly spot datasets that are missing gene names, lack spatial coordinates, or have their data split across multiple files instead of being neatly organized in a single h5ad object.

To do this, we’re planning to build a tool that can cross-check datasets by looking at their metadata and file formats. This will involve some web scraping to automatically collect and organize all the relevant metadata from GEO. The main output will be comprehensive metadata tables that summarize the key features and file details of each dataset. We’ll also include some visualization options—like bar plots, scatter plots, and filtered tables or SQL queries—to help us explore and filter the datasets more effectively. Once we’ve curated a solid list of promising datasets, we can then use existing tools like GEOquery to download and assess only those that are most likely to meet our needs.

I think the biggest challenge will be building a good web scraping module for GEO. I already have a pretty clear idea of several key variables we can extract and where to find them on GEO DataSets web site, but I have no experience with web scraping. That part will probably be the most difficult, and it’s something I’ll need to learn as we go.

There are some existing solutions out there, like Polly (Elucidata) and GEO2R, which are designed to offer a complete end-to-end analysis pipeline. However, they don’t really address this initial step of getting a clear overview of the data format availability for every queried dataset. Our focus is specifically on making this first stage of dataset evaluation much more transparent and efficient.

Benefit: 
With that tool, we could evaluate whether there are available datasets that are actually useful for us on public resources so we can run a pre-analysis on our hypotheses before spending money on new experiments.  

Recommended software: 
Python, R, JavaScript(?) 

Test data: 
We will be working with omics technologies from GEO (scRNA-seq, Spatial Transcriptomic, etc).

Do any of the proposed datasets contain PHI, sensitive information, or otherwise confidential data?  No

Title: 
Optimized HPC Resource Allocation for CryoEM Processing

Submitter:
Jose Miguel de la Rosa Trevin (Structural Biology)

Summary: 
This challenge tackles the problem of efficiently utilizing the HPC cluster resources, specifically for CryoEM data processing jobs. Many users, most of them with a Biology background, struggle with job submission parameters due to the complexity of CryoEM software. Users must know internal implementation details in several workflow steps and understand the existing hardware infrastructure (e.g., number of GPUs, cores, RAM) to select the correct parameters. The main goal of the challenge is to develop an intermediate layer that allows the specification of the cluster resources and automatically proposes the best parameters for specific jobs that typically run for hours or days. Developing a web portal to track utilization better and facilitate job submissions would be a beneficial side product. Addressing this challenge will require a collaborative team with experience in CryoEM data processing, software development, and HPC maintenance. 

Bernefit: 
Solving the proposed challenge will benefit final users, who can focus more on data analysis rather than struggling with job submission parameters. The use of the existing resources will be more efficient, reducing waiting times and accelerating the generation of results. Additionally, better tools will be available to monitor resource utilization, helping with more informed decisions about time allocation or future hardware acquisition. Although the initial scope of the challenge focuses on CryoEM, the solution could be a helpful starting point for other scientific areas facing similar problems. Similarly, the results could be extended to other HPC centers with different hardware configurations. 

Recommended software: 

• Programming languages: Python, Bash
• Web and visualization: HTML, JavaScript, Bootstrap, Matplotlib, Highcharts, Plotly, EMhub
• Database and files: SQLite, JSON, Redis
• HPC systems: Linux, SLURM, LSF, parallel filesystems, scratch
• Version control: Git

Do any of the proposed datasets contain PHI, sensitive information, or otherwise confidential data?  No

Title:
Pediatric CT Organ Segmentation: Building Smarter Model

Submitter:
Paul Yi (Radiology)

Summary:
Organ segmentation on pediatric CT scans is a foundational step for downstream applications in pediatric cancer care, including tumor response assessment, radiation treatment planning, and surgical navigation. However, existing state-of-the-art AI segmentation tools—like TotalSegmentator (https://totalsegmentator.com/)—are trained on adult datasets and perform poorly when applied to pediatric populations due to anatomical differences and data scarcity.

This challenge invites participants to take on the critical and exciting task of developing high-performing pediatric CT segmentation models using a small curated dataset of pediatric scans. The main obstacle is achieving generalizable model performance despite the limited volume of pediatric training data—a persistent bottleneck in pediatric AI.

Why are existing solutions inadequate? Current models (virtually all trained on adults) fail to generalize to pediatric anatomy, and there are virtually no open-source, pediatric-specific AI segmentation tools available. This is a massive gap in the field, especially given the high clinical impact of accurate organ segmentation in children with cancer.

Participants can adopt either data-centric approaches—such as leveraging public datasets, performing clever data augmentation, or using generative AI to synthesize pediatric data—or model-centric approaches, like adapting foundation models, transfer learning, or novel fine-tuning strategies. Or both! The goal: push the boundaries of what’s possible with limited pediatric data.

This challenge is ideal for creative hackers excited to bridge the gap in pediatric AI and build something that could be the seed for the first open-source, pediatric-specific organ segmentation toolkit. Let’s hack the data gap—and help kids in the process.

Benefit: 
Solutions developed through this challenge would lay the groundwork for the first open-source, pediatric-specific organ segmentation models—filling a critical gap in medical AI. Just as TotalSegmentator has transformed image analysis workflows in adult radiology by providing high-quality, off-the-shelf segmentation tools, an analogous resource for pediatrics would catalyze similar advances across pediatric oncology, surgery, and radiation therapy.

For our lab, these solutions would accelerate research in AI-driven precision imaging for childhood cancers, enabling faster prototyping, better quantitative biomarkers, and more consistent image analysis pipelines. For the broader scientific and clinical community, open access to pediatric-specific segmentation models would democratize pediatric imaging research, enable reproducibility, and reduce the barrier to entry for developing advanced imaging tools tailored to children—who are too often left behind in mainstream AI development.

By seeding an open-source solution, this challenge could unlock countless downstream applications, from improving clinical trial efficiency to powering large-scale imaging studies and enabling next-generation AI tools for pediatric care.

Recommended software:

• Programming Languages & Frameworks: Python, PyTorch, TensorFlow
• Segmentation Frameworks: nnU-Net, MONAI
• Foundation Models & Pretrained Networks: SAM (Segment Anything Model), MedSAM, BioMedCLIP, Swin UNETR
• Data Handling & Preprocessing: SimpleITK, NiBabel, NumPy, Pandas
• Visualization & Annotation: 3D Slicer, ITK-SNAP, napari
• Synthetic Data Generation: Diffusion models, generative adversarial networks (GANs), and domain adaptation tools
• Training Platforms: Google Colab, Jupyter, Docker, GitHub, Hugging Face

Test data:

• Pediatric CT Segmentation Dataset (TCIA): A curated collection of pediatric CT scans with organ segmentation labels—ideal for training and validating pediatric-specific models. https://www.cancerimagingarchive.net/collection/pediatric-ct-seg/ 
• CT-ORG Dataset (TCIA): A well-established dataset of abdominal CTs with organ labels, primarily from adult patients—useful for pretraining or transfer learning. https://www.cancerimagingarchive.net/collection/ct-org/ 
• TotalSegmentator Abdominal Organ Segmentation Dataset (Zenodo): A recent open-source release containing abdominal organ segmentations in CT, helpful for cross-domain experiments and data augmentation. https://zenodo.org/records/10047292 
• Medical Segmentation Decathlon: A diverse set of medical imaging segmentation tasks, including abdominal CTs and other modalities. http://medicaldecathlon.com/
• Other public medical CT datasets from sources such as TCIA, Synapse, and Zenodo may be used for data augmentation, pretraining, or domain adaptation. 

These datasets can support both pediatric-specific model evaluation and transfer learning strategies from adult imaging domains.

These tools will enable both data-centric and model-centric experimentation, from dataset curation to training and evaluating segmentation models.

Do any of the proposed datasets contain PHI, sensitive information, or otherwise confidential data?  No

Title:
AI-Driven Clinical Data Exploration Platform

Submitter:
Zhuo Qu (Biostatistics)

Summary:
Design a lightweight prototype platform that allows researchers to interact with de-identified clinical data (such as SEER or CCSS) using a local AI assistant. This platform should support variable exploration, filtering, data visualization, frequency table summary, regression analysis, and AI-guided code suggestions — all within a secure environment that respects data-sharing agreements and does not expose PHI.

Main Obstacle: 
Current clinical data sharing systems are siloed, manual, and rely heavily on back-and-forth between analysts and researchers. The main barriers include:

1. Lack of intuitive interfaces for exploring complex data dictionaries,
2. No AI guidance for variable selection or analysis code generation,
3. Compliance bottlenecks related to PHI exposure and DUA tracking.

Why Existing Solutions Are Inadequate: 
Most data portals are static, require IT overhead, and do not incorporate local LLMs due to privacy risks. Public tools like REDCap or SEER*Stat lack dynamic analysis guidance, chat capabilities, or regression modeling features tied to conversational input.

Feasibility in 72 Hours (Hackathon Scope): 
A small team (2–8 members) can focus on building an MVP prototype that includes:

I will prepare the pre-cleaned de-identified cancer survivorship datasets first: either Surveillance, Epidemiology, and End Results (SEER) or Childhood Cancer Survivorship Study Data (CCSS).

1. A web-based variable browser (React + D3.js),
2. Chat interface (e.g., with llama.cpp or GPT4All backend),
3. File upload for CSV + proposal PDFs,
4. Metadata parser for public data dictionaries (e.g., CCSS),
5. AI-suggested variable selection and simple regression code (Python/R).

Benefit:

For my biostat lab: 
It reduces a lot of time spent by analysts manually matching variables to requests, improves consistency in data handling and DUA compliance, and will achieve higher efficiency in data analysis pipeline through AI advice.

For the scientific community:

1. Promotes reproducible research in survivorship and oncology,
2. Reduces redundant analyst labor,
3. Encourages safe, AI-assisted data access at scale.

Recommended software:

• Frontend: React, Tailwind, D3.js, ShadCN/UI
• Backend: Node.js or Flask (Python), SQLite/PostgreSQL
• LLM: llama.cpp, GPT4All, Ollama
• Visualization: Recharts, Plotly, Pandas Profiling
• Statistical Tools: Statsmodels, Scikit-learn, R (via plumber API)
• Security: Local-first deployment, file encryption, audit logging

Test data:

• SEER Data (de-identified): Available publicly from SEER
• De-identified CCSS data with available dictionaries
• Data cited publications to train AI models to better recommend eligibility criteria, control variables, and statistical methods

Do any of the proposed datasets contain PHI, sensitive information, or otherwise confidential data?  No

Title:
Standardized and Simple Pipeline Execution via OnDemand

Submitter:
Jared Andrews (Developmental Neurobiology)

Summary:
St. Jude has limited options for easily run, best practices computational workflows easily accessible to anyone and backed by the computational resources available via the HPC. This results in tremendous redundant and unnecessary efforts to develop "pipelines" and creates obstacles for data harmonization across the institution.

nf-core (https://nf-co.re/) is a global community that provides open-source best practices Nextflow pipelines for most major data modalities. Notably, these pipelines are actively maintained and updated to adjust to new technologies, methods, and tools. They also utilize versioned releases with fixed software versions, ensuring reproducibility. These pipelines are already broadly used at St. Jude, though they require a small amount of initial effort to set up and properly configure, and the reference data used varies from group to group.

The recent launch of OnDemand at St. Jude has democratized access to many computational tools and environments, backed by the resources of the HPC and easily accessible through a web browser. This challenge proposes making the nf-core pipelines for common data modalities (RNAseq, ATACseq, WGS/WES) runnable via the OnDemand interface, allowing users to merely have to point to a samplesheet for their data. 

Benefit:
The ability to run these pipelines through OnDemand would provide broad access to easily runnable pipelines, properly configured for the St. Jude HPC, using consistent reference data. This would streamline data harmonization efforts, decrease the barrier to entry for generalized data processing, and pave the way for additional internal and external pipelines to be run through OnDemand.

It will save St. Jude researchers time, reducing the effort spent on establishment of well-defined best practice pipelines, and thus enabling more specialized and novel development work.

Recommended software:
Familiarity with running nextflow/nf-core pipelines, nextflow configurations, OnDemand, job templating

Test data:
Depends on the pipelines used, test data is readily found on GEO for all data modalities.

Do any of the proposed datasets contain PHI, sensitive information, or otherwise confidential data?  No

Title:
Patient-Specific Radiology Report Summarization with LLM

Submitter:
Andrew Smith (Radiology)

Summary:
The challenge is to develop and optimize a LLM to create patient friendly radiology reports from clinical radiology reports. The effort should be on translating the report conclusion to something that is patient friendly. We can share thousands of deidentified reports to assist with fine tuning a LLM. 

Benefit:
We will implement and test the best solution(s) first in prospective research and then in clinical practice. 

Recommended software:
Open source LLMs

Test data:
There may be limited public datasets. We can provide large datasets. 

Do any of the proposed datasets contain PHI, sensitive information, or otherwise confidential data? No

Title:
AI Chat Agent for Family Resource Navigation

Submitter:
Matthew Read (Collaboration Services)

Summary:
Utilize existing AI technologies to create a chat agent specifically trained on helping St Jude families navigate their way through treatment and care. It should incorporate wayfinding on campus, have knowledge about specific St Jude resources to direct users to, be able to provide age appropriate answers to questions regarding diagnosis and treatment, and be able to guide parents and guardians through documents like consent forms and care guides.

Benefit:
It would be a huge quality of life improvement for our patients and families to have access to an AI agent at their fingertips that can do everything from:

1. Answer their questions about disease and treatment, 
2. Help them navigate on campus and possibly even parts of the Memphis area, 
3. Break down complicated legal and medical language into something easier to understand,
4. Order transportation and room service,
5. Schedule and set reminders for appointments and events on campus they might be interested in.

Recommended software:
ChatGPT, Claude, Gemini, and others. 

Test data:
Not sure.

Do any of the proposed datasets contain PHI, sensitive information, or otherwise confidential data? I think the base agent wouldn't have access to personal information, but a patient or guardian could choose to give it access for personalized services.

Title:
Comprehensive PacBio Iso-Seq Analysis Pipeline

Submitter:
Zhongshan Cheng (Center for Applied Bioinformatics)

Summary:
Our proposed challenge is to develop an open-source pipeline that can:

1. Quantify differentially expressed genes and isoforms across multiple samples,
2. Detect differential isoform usage (DIU),
3. Identify potential fusion genes, and
4. Provide intuitive visualizations of isoform structures and fusion events.

This project is feasible for a small team (~3 people) to prototype within 72 hours, using conda for reproducibility and public datasets for benchmarking.

Benefit:
For Center for Applied Bioinformatics and Other St. Jude Researchers: As PacBio Iso-Seq becomes increasingly adopted at St. Jude and by external researchers, a reliable pipeline for analyzing and visualizing this data will be crucial. Many researchers now conduct experiments involving transcriptome profiling in cancer and rare diseases, where insights into isoform variation or fusion events can be biologically significant. Our center needs to help St. Jude researchers to analyze these PacBio Iso-Seq data with a fast pipeline that can accelerate hypothesis generation and reduce the burden of complex downstream analyses.

For the Broader Scientific Community: This workflow would provide a standardized, reproducible, and user-friendly approach to analyze PacBio Iso-Seq data, enabling broader accessibility to isoform-level insights. It will empower labs without deep computational expertise to uncover novel splicing and fusion patterns in diseases like cancer, neurodevelopmental disorders, or immune dysfunction.

Recommended software:
We plan to use the following tools and platforms:

• minimap2 – for long-read alignment
• TAGET – for accurate gene/isoform annotation and quantification
• SQANTI – similar to TAGET but for comparison results between the two software
• Conda – for environment management and reproducibility

The pipeline will be implemented to run on Linux system, and packaged for easy installation.

Test data:
We will use the publicly available Iso-Seq dataset from PacBio:

• Melanoma Iso-Seq dataset: https://downloads.pacbcloud.com/public/dataset/Melanoma2019_IsoSeq/

This dataset provides a rich source of tumor and normal transcriptomes suitable for demonstrating all key functionalities of the pipeline, including differential expression and fusion detection.

Do any of the proposed datasets contain PHI, sensitive information, or otherwise confidential data?  No.

Title:
Integrated Mitochondrial DNA Analysis Pipeline

Submitter:
Mondira Kundu (CMB)

Summary:
The challenge aims to integrate tools for mitochondrial DNA (mtDNA) analysis into a unified, streamlined workflow. The objective is to combine the existing WGS mtDNA pipeline, variant allele frequency (VAF) analysis tools, and MitoEdit into a deployable solution. This integrated workflow will enable researchers to efficiently analyze mtDNA sequencing data, identify variants suitable for editing, and predict optimal editing windows to facilitate the design of constructs for engineering specific mtDNA variants into cells.

Main Obstacle: The primary obstacle is the lack of a cohesive framework that combines these tools into a single, user-friendly pipeline. While the WGS mtDNA pipeline has already been implemented using Nextflow, downstream VAF analyses and MitoEdit remain separate, requiring manual intervention and computational expertise to link their outputs. This fragmentation reduces accessibility and scalability, particularly for researchers without bioinformatics experience.

Why Existing Solutions Are Inadequate:  Although individual tools for mtDNA analysis exist, they function independently and require significant effort to integrate their outputs manually. This fragmented approach is inefficient, prone to errors, and inaccessible to researchers unfamiliar with bioinformatics workflows. Existing solutions fail to provide an automated, seamless pipeline for mtDNA analysis and editing prediction, which is essential for advancing mitochondrial research and engineering.

Proposed Solution: We propose the development of an integrated Nextflow workflow that combines the WGS mtDNA pipeline, downstream VAF analysis, and MitoEdit into a unified framework. Nextflow’s modular and scalable design will facilitate efficient deployment and reproducibility, enabling researchers to analyze mtDNA data and predict editing windows without requiring extensive computational expertise.

Feasibility: With careful planning, a small team (2–8 people) can develop a prototype of the integrated workflow within 72 hours. The WGS mtDNA pipeline is already implemented in Nextflow, and the downstream VAF analyses are relatively straightforward to integrate. MitoEdit can be incorporated as an additional module within the workflow, ensuring compatibility and automation. This project is achievable within the timeframe and will significantly enhance accessibility, usability, and efficiency for mtDNA analyses.

Benefit:
Solutions to our proposed challenge would provide significant benefits to both our lab and the broader scientific community by coupling the analysis of mtDNA variants with the ability to engineer these variants, enabling deeper insights into mitochondrial biology and disease. The established workflow based on Nextflow can be easily shared with internal and external research groups, fostering collaboration and reducing redundant development efforts.

Benefits to Our Lab: Integrating the WGS mtDNA pipeline, variant allele frequency (VAF) analysis tools, and MitoEdit into a unified framework would streamline the identification of functionally impactful mtDNA variants and facilitate the design of constructs for engineering specific mutations into cells. This coupling of variant analysis with engineering capabilities would allow our lab to experimentally validate the functional consequences of mtDNA mutations and study their role in disease progression. By automating these processes, we could significantly reduce the time required for both computational analyses and experimental design, accelerating our ability to test hypotheses about the impact of mtDNA mutations on cellular metabolism and stress responses.

Benefits to the Greater Scientific Community: For the broader scientific community, the ability to link mtDNA variant analysis with targeted engineering would be transformative. This integrated workflow would enable researchers to efficiently identify variants of interest and design constructs to introduce these mutations into experimental models, bridging the gap between computational predictions and functional validation. Such capabilities are critical for studying mtDNA mutations in contexts like aging, cancer, and mitochondrial diseases, where understanding the functional impact of specific mutations is essential for developing therapeutic strategies.

By democratizing access to these tools, the workflow would empower researchers to investigate the evolutionary dynamics and cellular consequences of mtDNA mutations with unprecedented precision. Coupling analysis with engineering would also facilitate collaborative efforts to model disease-associated mutations, test therapeutic interventions, and uncover new insights into mitochondrial function and dysfunction. Ultimately, this solution would advance mitochondrial research across disciplines, accelerating discovery and innovation in both basic and translational science.

Recommended software:
Nextflow, R-dplyr, R-ggplot and Python-pandas.

Test data:
Platinum Genome, CCLE 

Do any of the proposed datasets contain PHI, sensitive information, or otherwise confidential data? No

Title:
Chemical Registration Database Search Interface

Submitter:
Nathaniel Twarog (Chemical Biology and Therapeutics)

Summary:
Our group, Lead Discovery Informatics, maintains a chemical registration database of over a million compounds used by dozens or hundreds of scientists at St. Jude every day.  This database contains structural information for nearly every one of these compounds; we would like to build a browser-based app in which researchers can query this database by chemical structure.  Ideally this tool would allow a user to input a chemical structure in one of several formats -- SMILES, ChemDraw format, or even drawn in the app itself -- and identify or rank compounds in the database based on exact matches, substructure matches, or structural similarity.  More advanced options might even include pharmacophore similarity, or even options to search larger existing virtual databases outside of the institution.  The team would need to develop the back-end of the app (including any data needed to be stored or cached in the database) as well as a front-end interface to perform searches of the database.

Benefit:
Drug naming and identification is notoriously ambiguous, so identifying a compound by its actual chemical structure can be absolutely essential when locating that compound in a registration database. Researchers investigating the chemical interaction of drug with a given protein may wish to locate other compounds possessing the drug's substructure believed to bind to that protein; in other circumstances, they may wish to evaluate compounds with similar chemical structures, similar chemical properties, or similar pharmacophore patterns, to see if those related compounds produce similar or contrasting behaviors.  The ability to search by structure will greatly enhance the research value of this massive information resource.

Recommended software:
The database in question is written and implemented in MySQL on a St. Jude Linux virtual machine.  Backend cheminformatics can likely be written in Python using an existing open-source cheminformatics library such as OpenBabel or RDKit.  The web application itself can be written in the web server languages of choice, with options including apache/PHP, Django, HTML, Javascript and CSS.

Test data:
A static test-version of the database can likely be prepared for use during the biohackathon, as well as a smaller test-version for prototyping.

Do any of the proposed datasets contain PHI, sensitive information, or otherwise confidential data?  No

Title:
Image-to-SMILES Conversion Tool

Submitter:
Vyoma Sheth (Structural Biology)

Summary: 
The accurate conversion of chemical structure images into SMILES (Simplified Molecular Input Line Entry System) is a critical task in cheminformatics, yet existing tools often face significant limitations in practical  setups. Current solutions are frequently plagued by issues such as low accuracy, poor usability, and restricted access due to paywalls, creating barriers for researchers who rely on these tools for chemical data extraction and analysis.

To address these challenges, the Center for Data Driven Discovery proposes a collaborative initiative at the upcoming St. Jude Biohackathon to develop innovative solutions for image-to-SMILES conversion. This initiative aims to create a tool that is accurate, user-friendly, and accessible to the broader scientific community.

Benefit:
Image-to-SMILES conversion enables the digitalization and sharing of chemical structures, streamlining data exchange and integration across platforms. It supports cheminformatics tasks like database searches, molecular modeling, and structural analysis. By automating structure digitization, it reduces manual work and enhances research efficiency. Additionally, it generates structured data for machine learning applications, aiding in molecular property prediction and accelerating drug discovery workflows. 

Recommended software:

• RDKit: Cheminformatics toolkit for SMILES handling, molecule visualization, and property calculations.
• OpenCV: Useful for pre-processing and enhancing chemical structure images.
• PyTorch: Deep learning frameworks for building and training image recognition models.
• ChemProp/ DeepChem: A machine learning library 
• Huggingface: A repository and framework with pre-trained image analysis deep-learning models.

Test data:

• STAKER: Contains 30,000 images derived from USPTO data with 256×256 px resolution. Molecules average 24 atoms, ranging from 7 to 51.
• USPTO: Includes 4,852 images (average resolution 649×417 px) from US patents, with molecules averaging 28 atoms (range: 10–96).
• UoB (University of Birmingham): Offers 5,716 images (762×412 px) featuring smaller molecules, averaging 13 atoms (range: 4–34).
• CLEF: Features 711 images from the CLEF test set with 1243×392 px resolution. Molecules average 26 atoms, ranging from 4 to 42.
• JPO (Japanese Patent Office): Comprises 365 images (607×373 px) with an average of 20 atoms (range: 5–43). This dataset presents additional challenges due to poor image quality, textual labels (including Japanese characters), and irregular features like variable line thickness. 
• CDD Public Access Data: Contains 3165011 compounds derived from 141 public datasets 
• ChEMBL: manually curated database of 2,496,335 bioactive molecules with drug-like properties

Do any of the proposed datasets contain PHI, sensitive information, or otherwise confidential data? No

Title:
Reusable Form Modules for Shared Resource Management

Submitter:
Susanna Downing/Melissa Mann (Shared Resources)

Summary:
The many shared resources across our organization frequently need to create similar forms in SRM2 to support common tasks such as equipment reservations, service requests, and sample submissions. However, the current process for building these forms is that a representative from the core meats with IS and then a completely custom form is created. 

Main Obstacle: Each core is independently developing its own forms, which leads to duplicated effort as they are also not currently aware of what each core is doing, user confusion due to inconsistent form design, and increased maintenance burden. There is no streamlined way to share or reuse components that are common across multiple resources.

Why Existing Solutions Are Inadequate: SRM2 lacks modular tools or templates that would allow shared resources to easily create and customize forms based on common needs. Without reusable modules, each resource must reinvent solutions, even for standard workflows. This limits efficiency, scalability, and user experience across the system.

Benefit:
Implementing reusable form modules in SRM2 would significantly reduce the time and effort required to create and maintain forms. It would also provide an easy starting template for new forms allowing time to focus on needed customization. This would lead to faster onboarding, more consistent user experiences, and fewer errors in request submissions.

For the broader St. Jude community, standardized and efficient workflows across cores would improve access to shared resources, reduce delays in service, and enhance reproducibility by ensuring consistent data collection and tracking. Overall, this would promote more seamless collaboration and higher-quality research outcomes.

Recommended software:
SRM2  (note I am not sure what they currently use for this and need to ask someone in IS)

Test data:
N/A

Do any of the proposed datasets contain PHI, sensitive information, or otherwise confidential data? No