thesis_implementation

Transfer single-cell foundation model representations to predict drug resistance in cancer

Code for my thesis evaluating bulk RNA-seq, single-cell (pseudobulk), and foundation-model embeddings (scGPT, scFoundation) for per-drug response prediction (LN(IC50)) on cancer cell lines.

TL;DR: Strong, well-regularized bulk baselines are hard to beat. Pseudobulk rarely surpasses matched bulk, and zero-shot transfer from single-cell foundation models does not consistently outperform classical approaches across drugs.

---

Repository structure

Plain-text (copy-safe) layout:

.

• bulk_state_of_the_art/ # Bulk RNA-seq pipelines, models, CV & evaluation
• envs/ # Conda environment YAMLs (one per used env)
• maps/ # ID maps / lookups (genes, models, tissues, etc.)
• scGPT/ # scGPT embedding extraction + downstream heads
• single_cell/ # scRNA-seq QC -> pseudobulk -> preprocessing
• .gitignore
• README.md

---

Setup

Create the Conda environments

This repo includes YAMLs under envs/. Create the ones you need (names may vary—use whatever is in envs/ on your branch):

# from repo root
ls envs

# create envs (examples; adjust to your files)
conda env create -f envs/base.yml
conda env create -f envs/scgpt.yml
conda env create -f envs/scfoundation.yml

# activate
conda activate <env-name-from-yaml>