St. Jude KIDS26 BioHackathon Challenges

Title: AI-Guided Discovery of Pediatric-Specific INO80/SWR1/SRCAP Chromatin Remodeler Subunits and Adaptors

Submitter: Vinesh Vinayachandran (Computational Biology) and Tian Cai (Structural Biology) 

Summary: This project develops an AI-guided framework to identify accessory subunits and adaptor proteins associated with the INO80/SWR1/SRCAP chromatin remodeling complexes that may represent pediatric cancer-specific dependencies. Participants will integrate co-expression, co-essentiality, functional dependency, and transcriptomic data to train predictive models that nominate novel candidate genes across the genome. The workflow emphasizes module-aware discovery of chromatin regulators that may be selectively required in pediatric malignancies. 
Benefit: The approach aims to uncover therapeutic vulnerabilities that are more selective than targeting broadly essential chromatin-remodeling core complexes. By prioritizing pediatric-specific dependencies, the project may identify candidates suitable for downstream functional validation and drug-discovery efforts. The resulting framework could be extended to other protein complexes and disease contexts. 

Tools: Python, Random Forest, XGBoost/LightGBM, logistic regression, Jupyter notebooks, DepMap, CCLE 

Test Data: DepMap CRISPR dependency data, CCLE RNA-seq data, TARGET/PedcBioPortal annotations, curated chromatin remodeler gene sets 

Any PHI, sensitive information, or otherwise confidential data use: No

Title: AI-Assisted Chat Interface for Clinical Data

Submitter: Sri Sivani Charan Yalamanchi (Biostatistics) and Sadie Mirzaei (Biostatistics) 

Summary: This project aims to build a conversational interface that enables researchers to query complex clinical datasets using natural language. A retrieval-augmented generation (RAG) architecture combined with Text-to-SQL translation will allow users to retrieve structured information and generate downloadable outputs without writing database queries. The prototype will focus on publicly available clinical datasets while establishing a framework that can be extended to institutional research resources. 

Benefit: The platform lowers barriers to data exploration by providing self-service access to structured clinical information. Researchers can iterate on hypotheses more quickly while reducing dependence on manual analyst-mediated workflows. The approach also demonstrates how natural language interfaces can improve accessibility of large clinical datasets while maintaining governance controls. 

Tools: Python, Flask, PostgreSQL, pgvector, RAG frameworks, LLM APIs, HTML/CSS/JavaScript 

Test Data: MIMIC III/IV datasets 

Any PHI, sensitive information, or otherwise confidential data use: No

Title: AI-Enabled Structuring of ARIA Guide Clinical Recommendations

Submitter: Monica Key (Global Pediatric Medicine), Mayuri Makan (Global Pediatric Medicine) and Darrell Gentry (Information Services) 

Summary: This project focuses on automatically transforming narrative clinical recommendations into structured, machine-readable formats suitable for downstream guideline management and digital implementation. Participants will develop an AI-assisted pipeline capable of extracting recommendation elements such as diagnoses, treatment guidance, risk stratification, and clinical notes from Word documents and mapping them into standardized templates. The solution will combine natural language processing with transparent validation workflows. 

Benefit: Automating guideline structuring can significantly reduce manual data-entry effort while improving consistency, traceability, and reproducibility. The resulting workflows could accelerate living-guideline maintenance and facilitate integration into decision-support systems and clinical information platforms. The project also provides a reusable model for converting narrative clinical content into computable resources. 

Tools: Python, python-docx, Pandas, OpenPyXL, Azure OpenAI or equivalent LLMs, Streamlit/Flask 

Test Data: ARIA Guide Word documents and Excel-based guideline resources 

Any PHI, sensitive information, or otherwise confidential data use: No

Title: AI/ML-Driven Sequencing Run Planner for NGS Core Operations

Submitter: Daniel Darnell (Development Neurobiology) 

Summary: This project seeks to apply machine learning and optimization methods to improve next-generation sequencing run planning and resource allocation within core sequencing operations. Participants will develop predictive and scheduling approaches that account for sample characteristics, instrument constraints, throughput objectives, and operational efficiency. The goal is to create a prototype decision-support system that assists with sequencing run composition and execution planning. 

Benefit: Improved run planning can increase instrument utilization, reduce operational bottlenecks, and improve turnaround times for sequencing services. The resulting framework may support more efficient allocation of sequencing resources while minimizing wasted capacity. These capabilities could ultimately benefit both core facilities and downstream research programs. 

Tools: AI/ML frameworks, scheduling and optimization tools, Python, data visualization frameworks 

Test Data: Sequencing operations and run-planning datasets 

Any PHI, sensitive information, or otherwise confidential data use: No

Title: AnTelOpe: An LLM-Powered Web Platform for Nextflow Pipeline Orchestration and Monitoring

Submitter: Haidong Yi (Center for Applied Bioinformatics) 

Summary: AnTelOpe integrates a conversational AI assistant directly into a web-based Nextflow workflow management platform. The system enables users to configure, launch, monitor, and troubleshoot bioinformatics pipelines through natural language interactions while leveraging live access to workflow metadata and execution logs. The prototype focuses on reducing technical barriers associated with large-scale computational analyses. 

Benefit: Researchers can manage sophisticated genomics workflows without requiring extensive command-line expertise. Automated interpretation of runtime errors and job status information can substantially reduce troubleshooting effort while improving accessibility of shared computational resources. The architecture also provides a governance-friendly alternative to cloud-only workflow assistants. 

Tools: Nextflow, nf-core, OpenAI models, object storage systems, web application frameworks 

Test Data: nf-core parameter schemas, nf-core test datasets, synthetic FASTQ files, public reference genomes 

Any PHI, sensitive information, or otherwise confidential data use: No

Title: AutoCAB: Turning Everyday Scientific Workflows into Reusable AI Agent Skills

Submitter: Niu Beifang (Center for Applied Bioinformatics) 

Summary: AutoCAB explores how recurring scientific workflows can be automatically transformed into reusable AI-agent skills. The project seeks to identify repeated analyst activities, compare them against existing skill libraries, and generate structured skill definitions that can be reviewed, validated, and contributed back to a shared repository. The prototype emphasizes governance, explainability, and privacy-aware automation. 

Benefit: Converting repetitive analytical tasks into reusable skills can improve reproducibility, accelerate onboarding, and reduce duplicated effort across research teams. The resulting framework provides a pathway for institutional knowledge capture and structured AI-assisted workflow development. The project also contributes toward building more transparent and maintainable AI automation ecosystems. 

Tools: Python, sentence-transformers, scikit-learn, FastAPI/Streamlit, local LLMs, GitHub 

Test Data: Public repositories, synthetic workflow traces, public skill libraries 

Any PHI, sensitive information, or otherwise confidential data use: No

Title: CLEAR: Clinician-Led, Easy-to-read, Accessible Reaffirmed Instructions

Submitter: Matt Davis (Hospitalist Program) 

Summary: This project focuses on improving the readability, accessibility, and consistency of patient-facing clinical instructions using AI-assisted language transformation and validation. The goal is to generate clinically accurate instructions that are easier for patients and families to understand while retaining essential medical information. The workflow will support transparent review and clinician oversight. 

Benefit: Improved communication can promote health literacy, increase adherence, and reduce confusion around treatment plans and care instructions. Standardized processes for instruction development may also improve consistency across clinical programs. The resulting tools could support scalable creation and maintenance of patient-centered educational materials. 

Tools: NLP frameworks, large language models, readability assessment tools, web interfaces

Test Data: Clinical instruction content and synthetic examples 

Any PHI, sensitive information, or otherwise confidential data use: No

Title: Counting every cell in the human brain: 3D localization of cell nuclei in a human brain  

Submitter: Caitlin Freeman (Imaging Sciences)  

Summary: This project aims to develop computational methods for large-scale three-dimensional localization and quantification of cell nuclei across human brain tissue volumes. Participants will work with imaging-derived datasets to identify, segment, and spatially map nuclei at scale while addressing challenges related to volumetric image processing and cellular density estimation. The resulting workflow will generate spatially resolved cellular maps suitable for downstream neuroanatomical analysis. 

Benefit: Accurate whole-brain cell localization provides a foundation for studying brain architecture, cellular organization, and disease-associated spatial patterns. Scalable approaches for nuclei detection can support quantitative neurobiology studies and generation of cellular atlases. The project may also establish workflows applicable to other large-scale imaging datasets. 

Tools: Image analysis software, machine learning frameworks, 3D image processing tools 

Test Data: Human brain imaging datasets 

Any PHI, sensitive information, or otherwise confidential data use: No

Title: CureCheck: An R Shiny Application for Evaluating Cure Model Appropriateness  

Submitter: Durbadal Ghosh (Biostatistics)  

Summary: CureCheck is an interactive application that helps researchers determine whether statistical cure models are appropriate for a given oncology dataset. The platform implements RECeUS-AIC methodology, model comparison workflows, diagnostic visualizations, and sensitivity analyses to evaluate cure fractions and long-term survival assumptions. The goal is to make advanced survival modeling methods more accessible to biostatisticians and clinical investigators. 

Benefit: Appropriate use of cure models is increasingly important in pediatric oncology, where many patients become long-term survivors. By providing standardized assessments and reproducible workflows, the application can improve statistical rigor and reduce the risk of model misapplication. The project also fills a methodological gap by translating recent research advances into an accessible software tool. 

Tools: R Shiny, RECeUS-AIC, GitHub, shinyapps.io  

Test Data: Publicly available oncology survival datasets from R survival packages 

Any PHI, sensitive information, or otherwise confidential data use: No

Title: Establishing a Sample-Based Quantum Diagonalization (SQD) Workflow for Iron-Sulfur Cluster Electronic Structure  

Submitter: Amandeep Bhatia (Structural Biology)  

Summary: This project develops a hybrid quantum-classical workflow for modeling the electronic structure of iron-sulfur clusters, which are challenging systems characterized by strong electron correlation and complex magnetic interactions. Participants will implement Sample-Based Quantum Diagonalization (SQD) methods on simplified Fe-S model systems and compare results against classical quantum chemistry approaches. The prototype will emphasize reproducibility, scalability, and rigorous evaluation of low-energy electronic states. 

Benefit: Improved computational approaches for metalloproteins could enhance mechanistic understanding of biologically important cofactors involved in metabolism, DNA repair, and catalysis. The project serves as an entry point for applying emerging quantum computing approaches to biomedical research problems. It may also provide a reusable framework for future molecular simulation and drug discovery applications. 

Tools: Python, PySCF, Qiskit Nature, Qiskit Aer, qiskit-addon-sqd, ffsim 

Test Data: Publicly available molecular structure and Hamiltonian data for simplified Fe-S systems 

Any PHI, sensitive information, or otherwise confidential data use: No  

Title: FAIR Harmonization and Automated LabKey Integration of Legacy Sequencing Metadata  

Submitter: Cody Ramirez (Developmental Neurobiology)  

Summary: This project focuses on automated harmonization, validation, and standardization of heterogeneous sequencing metadata accumulated across years of research projects. Participants will build workflows to parse legacy metadata sources, identify inconsistencies, apply controlled vocabularies, and generate LabKey-compatible outputs. The goal is to improve interoperability and FAIR stewardship of historical sequencing datasets. 

Benefit: Automating metadata harmonization can reduce manual curation burden, improve reproducibility, and facilitate integration of historical datasets into centralized data management systems. The resulting framework supports institutional data-sharing initiatives and enables more efficient cross-study analyses. The methodology is broadly applicable across multiple omics technologies and research programs. 

Tools: Python, R, Nextflow, Docker, LabKey, ontology resources, GitHub 
Test Data: Internal sequencing metadata files, manifests, spreadsheets, and de-identified sequencing project records 

Any PHI, sensitive information, or otherwise confidential data use: Not specified  

Title: Genomic Scars Pipeline: Assessment and Visualization Shiny  

Submitter: Evan Savage (Chemical Biology and Therapeutics)  

Summary: This project will create a platform for assessing, visualizing, and interpreting genomic scars derived from pediatric whole-genome sequencing data. The application will integrate multiple genomic scar detection methods and provide interactive exploration of mutational signatures, structural variants, copy number changes, and DNA damage repair deficiencies. AI-assisted interpretation tools will help users investigate complex genomic patterns. 

Benefit: The platform will provide researchers with a unified framework for exploring genomic instability and DNA damage repair signatures in pediatric cancer. Interactive analysis capabilities can accelerate hypothesis generation and support translational studies seeking clinically actionable biomarkers. The approach also addresses an unmet need for integrated genomic scar assessment workflows. 

Tools: R Shiny, SigProfiler Suite, MuSiCAL, MutationalSignatures, LLM integration

Test Data: De-identified intermediate WGS-derived genomic scar datasets   

Any PHI, sensitive information, or otherwise confidential data use: No

Title: GRIN3D: From Linear Sequence to 3D Protein Hotspots  

Submitter: Abdelrahman Elsayed (Biostatistics)  

Summary: GRIN3D extends the GRIN statistical framework from one-dimensional genomic coordinates into three-dimensional protein space. Participants will integrate genomic lesion data with AlphaFold-derived protein structures to detect spatially clustered functional hotspots that may not be apparent through sequence-based analyses alone. The framework will support diverse genomic alterations, including mutations, indels, and structural variants. 

Benefit: Identifying spatially convergent lesions within protein structures can reveal functionally important regions involved in disease biology and therapeutic targeting. The project bridges genomics, structural biology, and statistical modeling to generate more biologically meaningful interpretations of cancer-associated alterations. The resulting methodology may be broadly applicable across disease areas. 

Tools: R, AlphaFold database, UniProt, statistical clustering approaches, visualization tools 

Test Data: Public pediatric T-ALL genomic datasets and AlphaFold protein structures   

Any PHI, sensitive information, or otherwise confidential data use: No

Title: Integrating Molecular and Clinical Trial Data to Predict AML Patient Outcomes for Novel Therapies  

Submitter: Nobel Makonnen (St. Jude Graduate School and Biostatistics)

Summary: This project develops methodologies for integrating molecular response profiles from preclinical studies with clinical trial prognostic models to estimate the potential impact of novel therapies in acute myeloid leukemia. Participants will automate extraction and integration of GEO molecular response data with TARGET-derived prognostic frameworks. The resulting workflow will generate predicted survival outcomes for hypothetical treatment combinations. 

Benefit: By prioritizing therapies predicted to produce the greatest survival improvements, the framework may accelerate translational research and guide experimental validation efforts. The methodology provides a data-driven approach for identifying promising therapeutic combinations before costly laboratory or clinical studies. The resulting computational workflows could be adapted to additional cancer types.  

Tools: R, Python, GEO, TARGET 

Test Data: Public GEO molecular response datasets and TARGET clinical trial datasets 

Any PHI, sensitive information, or otherwise confidential data use: No

Title: Multimodal Sleep Disordered Breathing (SDB) Event Classification  

Submitter: Lu Xie (Biostatistics) and Miguel Navarrete (Psychology and Behavioral Sciences) 

Summary: This project will develop machine learning models for automated detection and classification of sleep apnea and hypopnea events using multimodal physiological signals collected from home sleep testing systems. Participants will integrate respiratory, oxygen saturation, movement, and EEG-derived features to generate time-stamped event annotations and clinically relevant summary metrics. The workflow will be designed to perform robustly in noisy real-world environments.  

Benefit: Automated scoring has the potential to reduce clinician workload, improve consistency of sleep-study interpretation, and accelerate diagnosis of sleep-related breathing disorders. The project demonstrates how multimodal machine learning approaches can improve interpretation of physiologic monitoring data acquired outside controlled laboratory settings. The resulting framework may support more scalable and accessible sleep diagnostics. 

Tools: Python, MATLAB, ML classifiers, EDF-processing tools, visualization dashboards  

Test Data: Public NSRR datasets and de-identified sleep recordings  

Any PHI, sensitive information, or otherwise confidential data use: No

Title: Resource-Aware Multi-Project Workflow Orchestrator for Scalable Single-Cell Analysis

Submitter: Antonia Chroni (Development Neurobiology) 

Summary: This project proposes a workflow orchestration framework capable of coordinating multiple concurrent single-cell analysis projects while accounting for computational resource availability and scheduling constraints. Participants will develop intelligent job-allocation and execution strategies that maximize throughput across shared infrastructure. The focus is on scalable, reproducible management of computationally intensive omics workflows. 

Benefit: Improved workflow orchestration can increase infrastructure utilization, reduce queue bottlenecks, and improve reproducibility across large-scale single-cell analysis projects. The framework could support institutional demand for increasingly complex computational pipelines while enabling more efficient prioritization of shared resources. It also provides a foundation for scalable research computing operations. 

Tools: Workflow managers, orchestration frameworks, cloud/HPC resource schedulers 

Test Data: Single-cell analysis workflows and computational resource metrics 

Any PHI, sensitive information, or otherwise confidential data use: No

Title: SCOGS-Scribe: Automated Severity Grading for Sickle Cell Disease from Clinical Notes

Submitter: Joe Wardell (Global Pediatric Medicine) 

Summary: SCOGS-Scribe aims to automate application of the Sickle Cell Outcome Grading System (SCOGS) by converting unstructured clinical notes into standardized severity assessments. Participants will build a machine-readable representation of the SCOGS rubric, generate synthetic training data, and develop language-model-based grading workflows with transparent rationale generation. Interactive dashboards will visualize patient- and cohort-level severity patterns. 

Benefit: Automating SCOGS grading has the potential to dramatically reduce manual chart review effort while improving consistency and reproducibility across research studies and clinical programs. The resulting tools could accelerate outcomes research, clinical trial readiness, and longitudinal disease severity monitoring. The project also contributes toward creation of computable clinical standards for sickle cell disease. 

Tools: R, Shiny, LLMs, JSON/YAML, tidyverse, yardstick 

Test Data: Synthetic sickle cell disease clinical notes and public references 

Any PHI, sensitive information, or otherwise confidential data use: No

Title: St. Jude AI and Data Learning Assistant: Personalized Guidance Using The Turing Way and Institutional Knowledge

Submitter: Ty Michael (Computational Biology) and Saikat Nandi (St. Jude Graduate School)  

Summary: This project will build an AI-powered learning assistant that helps users discover training resources related to data science, AI, software development, and reproducible research. The assistant will combine content from The Turing Way with curated institutional resources and deliver role-aware recommendations through a conversational interface. The system is designed to support learners across a wide range of expertise and professional backgrounds. 

Benefit: Personalized navigation of educational resources can improve onboarding, knowledge discovery, and AI literacy across the institution. By reducing barriers to finding relevant guidance, the platform supports self-directed learning and broader adoption of data science best practices. The framework also demonstrates how open educational resources can be integrated with institutional knowledge. 

Tools: LLM, Python, vector databases, GitHub, retrieval-augmented generation frameworks (MyGPT) 

Test Data: The Turing Way, public educational resources, approved institutional materials 

Any PHI, sensitive information, or otherwise confidential data use: No

Title: St. Jude AI and Data Learning Assistant: Personalized Guidance Using The Turing Way and Institutional Knowledge 

Submitter: Ty Michael (Data Sciences) 

Summary: Researchers often struggle to identify relevant training resources in AI and data science due to fragmented materials across platforms. This project develops an AI-powered assistant that provides personalized, role-based guidance using curated internal and external knowledge sources.  

Benefit: It accelerates onboarding and skill development across the institution. It also promotes adoption of best practices in data science and open science.  

Tools: Retrieval-based AI systems 

Test Data: Knowledge repositories 

Any PHI, sensitive information, or otherwise confidential data use: Not specified 

Title: Cell Identity in Spatial OMICs Imaging Platforms: Cluster-Based vs Cell-Based Annotation

Submitter: Maycon Marção (Developmental Neurobiology) 

Summary: This project investigates two commonly used approaches for assigning cell identities in imaging-based spatial transcriptomics datasets: cluster-based and rule-based annotation. Participants will compare methodologies across Xenium and CosMx datasets, quantify agreement and discrepancies, and build reproducible analytical workflows for systematic evaluation. The objective is to better understand how annotation choices influence biological interpretation. 

Benefit: Improving understanding of annotation strategies can strengthen reproducibility and confidence in spatial omics analyses. The work may identify conditions where different approaches capture complementary biological information and establish best practices for future studies. The resulting workflows can support broader adoption of spatial transcriptomics technologies. 

Tools: Python, Scanpy, Squidpy, SpatialData, CellxGene, PanglaoDB 

Test Data: Public Xenium and CosMx spatial transcriptomics datasets 

Any PHI, sensitive information, or otherwise confidential data use: No

Title: QC Compass: Interactive Data Quality Guidance for Clinical Research Platforms

Submitter: Yichen Chen (Global Pediatric Medicine) 

Summary: QC Compass is an interactive platform that guides users through standardized data quality management workflows for clinical research systems such as REDCap, REDCap Cloud, and TrialMaster. Participants will develop workflow decision trees, database-specific guidance, query-management examples, and AI-assisted templates that support consistent quality control practices. The project emphasizes accessibility, reproducibility, and onboarding support. 

Benefit: Centralizing data quality management knowledge can improve consistency across studies, reduce training burden, and support reproducible research operations. The platform provides a scalable foundation for institutional quality-control guidance while helping researchers navigate complex electronic data capture ecosystems. It also promotes cross-team standardization of clinical research workflows. 

Tools: R Shiny, Python, HTML/CSS/JavaScript, LLM APIs, REDCap and TrialMaster documentation 

Test Data: De-identified QC workflows, query reports, synthetic datasets, workflow documentation 

Any PHI, sensitive information, or otherwise confidential data use: Yes (deidentified example data)