A Reproducible Baseline for Multi-Omics Drug-Response Prediction

Goal. Provide a clean, end-to-end baseline pipeline and API scaffold for predicting drug response from multi-omics data, with explicit stress-tests for missing modalities.

Why this repo

• Most “SOTA” papers are hard to reproduce. This is a usable starting point others can extend.
• Focus on robustness: how badly do models degrade as modalities go missing, and which simple strategies help?

What’s included

• Minimal FastAPI service (/predict) and Dockerfile
• Experiment matrix (configs/experiments.yaml) with missingness regimes (MCAR, block-missing)
• Baseline models (logistic, random forest, simple NN) — to be filled in the src/pipeline/ modules
• Unit tests and CI (GitHub Actions)
• Makefile with reproducible environment setup (pip, conda, Docker)

Quickstart

make env         # create/update conda env (auto-generates environment.yml)
make run         # run baseline pipeline
# or via Docker
make docker-build-train && make docker-run-train

Example request:

curl -X POST http://localhost:8000/predict -H "Content-Type: application/json"   -d '{"features":{"g1":1.0,"g2":-2.0,"g3":0.5},"masks":{"g1":1,"g2":1,"g3":0}}'

Environment Setup

Option A: Pip / Colab (lightweight)

pip install -r requirements.txt
python scripts/train.py --config configs/experiments.yaml

Option B: Conda + Makefile (recommended)

make env          # create/update env (pip inside conda)
make env-update   # update from environment.yml
make run          # run baseline
make freeze       # export pinned deps to requirements-lock.txt
make clean        # remove caches

Optional Conda extras (CUDA, MKL):

make env-conda CONDA_DEPS="pytorch pytorch-cuda=12.1" CONDA_CHANNELS="pytorch nvidia conda-forge"

Pure Conda convenience (no YAML/pip):

make conda-env

Option C: Docker (full reproducibility)

make docker-build-train
make docker-run-train
# after adding FastAPI app/main.py:
make docker-build-api
make docker-run-api

Download & Ingest Data

1. Download TCGA data:

   make data-tcga

   Downloads into:

   • data/raw/tcga/tcga_RSEM_gene_tpm.gz
   • data/raw/supplemental/Survival_SupplementalTable_S1_20171025_xena_sp

2. Ingest and preprocess:

   make data-real

   Extracts .tsv, skips duplicate copies, and organizes:

   data/raw/
     ├── tcga/
     │   ├── tcga_RSEM_gene_tpm.gz
     │   └── tcga_RSEM_gene_tpm.tsv
     └── supplemental/
         └── Survival_SupplementalTable_S1_20171025_xena_sp
   

3. Full reset (clean and reload end-to-end):

   make data-reset
   

Roadmap (execution order)

1. Phase 1: Baselines under complete vs missing data (MCAR, block-missing). Metrics: AUROC/AUPRC.
2. Phase 2: Simple robustness methods: mean/zero + mask, matrix completion (softImpute), cross-modal ridge (RNA→pseudo-proteome).
3. Phase 3: Productionization: clean configs, reproducible runs, API/dashboard.
4. Phase 4: Critical analysis vs frontier methods (GNNs, Transformers, CODE-AE, diffusion) — discussion only in the public version.

Discussion pointers (for the blog post)

• Missing-modality robustness is often under-reported. Publish curves of performance vs % missing.
• Advanced architectures add value only if they respect masks / domain shift; otherwise they overfit.

Contributing / Extending

• Drop in alternative models in src/pipeline/
• Add new missingness regimes in configs/experiments.yaml
• PRs for better baselines welcome