Project Theme

Assign Functions to Uncharacterized Genes

Project Summary

A significant portion of genes in bacterial and viral pathogens remain uncharacterized or annotated only as “hypothetical proteins.” This project uses AI and machine learning to predict functions for these genes by integrating multiple data types including sequence, structure, genomic context, and expression patterns. By combining deep learning on protein sequences, AlphaFold structure predictions, and comparative genomics, we aim to provide functional annotations with confidence scores, helping researchers understand pathogen biology and identify novel therapeutic targets.

Goals and Objectives

1. Develop multi-modal AI models for function prediction integrating sequence, structure, and context
2. Generate functional predictions for hypothetical proteins in priority pathogens
3. Provide confidence scores and supporting evidence for predictions
4. Create visualization tools for exploring predictions and evidence

Approach

Methods and AI/ML Approaches:

• Deep learning on protein sequences (transformers, CNNs)
• Structure-based function prediction using AlphaFold models
• Genomic context analysis (operons, gene neighborhoods, synteny)
• Transfer learning from well-characterized model organisms
• Ensemble methods combining multiple prediction approaches
• Graph neural networks for protein-protein interaction prediction

Implementation Steps:

1. Compile datasets of characterized proteins for training
2. Extract features: sequences, structures, genomic context
3. Train and validate multi-modal prediction models
4. Generate predictions for hypothetical proteins in BRC
5. Implement confidence scoring and evidence aggregation
6. Build visualization interface for predictions
7. Validate predictions with experimental data where available

Data and Resources Required

Resource Type	Source / Link	Description / Purpose
Data	BV-BRC genome annotations	Target genes for prediction
Data	UniProt, InterPro, Pfam	Characterized protein functions
Data	AlphaFold Database	Protein structure predictions
Data	STRING, BioGRID	Protein interaction networks
LLMs / AI Models	ESM-2, ProtTrans	Protein language models
Tools / Services	PyTorch, TensorFlow	Deep learning frameworks
Compute / Storage	GPU for model training	Large-scale predictions

Expected Outcomes / Deliverables

• Functional predictions for hypothetical proteins in priority pathogens
• Confidence metrics and evidence summaries for each prediction
• Trained models for function prediction
• Web interface or API for batch predictions
• Comparison with existing tools (InterProScan, BLAST, HHpred)
• Public repository with code, models, and documentation
• Publication-ready validation results

Potential Impact and Next Steps

Impact on:

• Infectious disease research: Accelerates functional annotation of pathogen genomes
• AI/ML automation: Advances protein function prediction methods
• Public health preparedness: Enables rapid characterization of emerging pathogens

Next Steps After Codeathon:

• Apply to all BRC genomes genome-wide
• Integrate with BRC annotation pipelines
• Add experimental validation suggestions
• Develop active learning for continuous improvement
• Create educational materials on protein function prediction

Technical Support Needed

• BRC genome data access
• AlphaFold structure predictions
• GPU compute for training
• Curated protein function datasets
• Mentor support from structural biologists

Team Information

Teams will be formed during the Codeathon. Ideal team composition:

• ML/AI Researcher: Model development and training
• Structural Biologist: Structure-function relationships
• Comparative Genomicist: Genomic context analysis
• Bioinformatician: Data integration and validation
• Software Developer: Interface and deployment