biomedical ai / spatial transcriptomics
TumorScope
an AI copilot for exploring tumor ecosystems from public spatial transcriptomics data.
overview
TumorScope is an AI-powered research demo for exploring tumor ecosystems from public single-cell and spatial omics data. The project loads a public 10x Genomics Visium breast cancer spatial transcriptomics dataset and runs an end-to-end analysis pipeline using Scanpy and Squidpy, including preprocessing, clustering, UMAP visualization, spatial tissue mapping, marker gene ranking, and tumor microenvironment program scoring.
The system scores rough spot-level biological programs such as tumor epithelial, T cells, cytotoxic T/NK, B cells/plasma, myeloid/macrophage, fibroblast/stromal, endothelial, and immunosuppression/exhaustion signatures. It also computes lightweight ligand-receptor communication hypotheses, helping surface possible tumor-immune and tumor-stromal signaling axes such as MIF-CD74, CXCL12-CXCR4, and TGFB-TGFBR.
To make the analysis easier to interpret, TumorScope generates structured evidence cards from plots, tables, marker programs, and communication signals. A local RAG-style AI copilot then answers biological questions using those generated evidence cards and curated literature snippets, while clearly separating hypothesis-generating insights from validated biological claims.
why it matters
TumorScope is my strongest biomedical AI research demo. it frames tumors as ecosystems of malignant, immune, stromal, vascular, and signaling programs, then connects spatial gene expression, marker programs, communication signals, and AI-grounded explanations into one workflow for careful hypothesis generation.
what i built
- built an end-to-end spatial transcriptomics analysis app for public breast cancer Visium data.
- implemented Scanpy preprocessing, clustering, UMAP visualization, spatial plotting, and marker gene ranking.
- added marker-based tumor microenvironment program scoring across epithelial, immune, stromal, vascular, and immunosuppressive signatures.
- built a lightweight ligand-receptor scoring module to identify hypothesis-generating tumor-immune communication signals.
- designed a local RAG-style AI copilot that answers questions from generated evidence cards and curated literature snippets.
- added scientific caution layers to avoid overclaiming from transcript-based spatial signals.