Deep Learning for Lithium Response Prediction in Bipolar Disorder

This project outlines a proposed hybrid study to develop and validate a deep learning pharmacogenomics (PGx) model that predicts lithium response in individuals with bipolar disorder (BD). The methodology integrates genomic, transcriptomic, methylation, and clinical data, using both retrospective datasets and a newly recruited prospective cohort.

Project Highlights

• Combines data from ConLiGen, PGBD, and 100 new BD patients
• Multi-omic inputs: SNPs, RNA-seq, DNA methylation, and clinical scales
• Late-fusion neural network with interpretable outputs via SHAP
• Emphasis on ethics, bias mitigation, and Canadian-aligned data governance

Included Files

• methodology/rewritten_methodology_by_anushka.md: Full study design (937 words)
• figures/Figures_draft3.pdf: Contains Figure 1 (Timeline Schematic), Figure 2 (Model Design), and Table 1 (Methodology Summary)

Future Directions

• Code implementation of the model (Python/PyTorch)
• Synthetic data mockup for public demonstration
• External validation using held-out retrospective subsets

Ethics & Governance

• Designed for Canadian REB approval under TCPS 2
• PIPEDA-aligned privacy protocols
• Equity-focused recruitment and fairness-aware modeling

“This project reflects the intersection of mental health, personalized medicine, and ethical AI-driven discovery.”