Sally332

👩‍🔬 Computational Biologist | Cancer Genomics & Interpretable AI

Building open, reproducible frameworks for multi-omic data integration, spatial transcriptomics, and computational pathology.

🔬 Research Vision

My work focuses on developing interpretable and reproducible AI frameworks for cancer genomics, uniting biological prior knowledge with multi-omic and spatial data. The central goal is to replace black-box prediction with mechanistic understanding—models that not only perform well but explain how genomic alterations, perturbations, and drug responses reshape cellular states. Each framework in this series emphasizes pathway- and network-level interpretability, cross-dataset generalization, and transparent benchmarking, establishing reproducible standards for computational oncology. Through this approach, I aim to bridge machine learning, systems biology, and translational research, advancing models that predict, explain, and validate biological mechanisms.

📂 Key Projects

The following key projects are part of the MM-KPNN framework family, a unified effort to develop concept-bottleneck AI models that embed biological knowledge directly into network architecture—ensuring interpretability, reproducibility, and mechanistic insight across multi-omic and spatial data.

1. SpatialMMKPNN

A modular and interpretable graph framework for spatial transcriptomics in the tumor microenvironment.

Combines Graph Attention Networks (GAT) with knowledge-primed decoding
Explains immune exclusion, stromal remodeling, and therapy-induced rewiring
Outputs attention maps, pathway overlays, and ligand–receptor driver rankings

2. MM-KPNN

Interpretable multimodal neural network integrating scRNA-seq + scATAC-seq using biological priors.

Decoder constrained by pathway and TF nodes
Provides mechanistic attributions at the pathway and regulator levels
A reproducible framework for multimodal interpretability and benchmarking

3. DrugResponse-GNN

Pathway-bottleneck graph neural network for drug-sensitivity prediction across pharmacogenomic panels.

Integrates multi-omic features, drug descriptors, and prior knowledge graphs
Focuses on cross-panel generalization (e.g., CCLE → GDSC)
Provides pathway-level interpretability and reproducible benchmarking

4. Perturbation-MMKPNN

Extends MM-KPNN to model drug and CRISPR perturbation responses at single-cell resolution.

Implements pathway and TF bottlenecks for interpretability
Measures attribution stability and supports counterfactual pathway editing
Designed for robust, cross-dataset perturbation benchmarking

Additional Repositories

Organoid_Analysis

A modular framework for computational analysis of organoid systems.

Addresses reproducibility, heterogeneity, fidelity, integration, and prediction
Integrates RNA and protein modalities with interpretable ML
Demonstrates end-to-end reproducibility through documented, result-embedded notebooks

Spatial_Transcriptomic_Mapping

Spatial mapping of tumor and metastatic architecture using 10x Visium transcriptomics.

Integrates curated gene programs to define epithelial, immune, stromal, and proliferative regions
Reveals spatial organization and regional heterogeneity across breast tumors and lymph node metastases
Fully documented, end-to-end notebook with embedded results and biological interpretation