Training Cancer Drug-Response Models in TensorFlow

The base predictor — no ensembling, no intervals. This is the model the later work combines.

_{Ved Piyush · PhD in Statistics · University of Nebraska–Lincoln}

---

What it predicts. Given a cancer cell line and a drug, how sensitive is that cell line to that drug — reported as IC50, the drug concentration needed to halve cell growth. Lower means more sensitive.

How. A graph convolutional network reads the drug's chemical structure as a graph of atoms and bonds, while the cell line's molecular profile enters as separate inputs. This follows the published DeepCDR architecture.

This repository contains only the base model. It performs no ensembling and produces no prediction intervals — that work lives in the companion repositories.

How it fits together

flowchart LR
    A["Drug structure<br/>as a graph"] --> C["Graph convolutional<br/>network"]
    B["Cell-line<br/>molecular profile"] --> C
    C --> D["Predicted IC50"]
    E["GDSC · CCLE · CTRPv2"] -.->|"transfer<br/>learning"| C
    style C fill:#0b7a64,color:#fff,stroke:#0b7a64

Loading

What the code does

Training runs across four public drug-screening datasets: GDSC-1 and GDSC-2 (Genomics of Drug Sensitivity in Cancer), CCLE (Cancer Cell Line Encyclopedia) and CTRPv2 (Cancer Therapeutics Response Portal).

Two training regimes are compared: learning from scratch (From_Scratch_Learning_on_GDSC_2.py) and transfer learning, where a model trained on one dataset is adapted to another (Transfer_Learning_on_GDSC_2). Several notebooks keep dropout active during training in a way that supports later uncertainty estimation (train_deepcdr_gcn_dropout_in_train*).

Performance is reported as Spearman correlation, R² and mean squared error. Data loading goes through improve_utils.py, from Argonne National Laboratory's IMPROVE framework for benchmarking drug-response models.

Running it

1. get_sample_data.ipynb pulls a small sample to check the data path works.
2. train_deepcdr_gcn*.ipynb trains the graph convolutional model.
3. Transfer_Learning_on_GDSC_2 adapts a trained model to a second dataset.
4. Get_Joint_Learner_GDSC_*_Preds.ipynb writes out predictions for downstream use.

Where to look first

• train_deepcdr_gcn.ipynb — the base model, trained from scratch
• Transfer_Learning_on_GDSC_2.ipynb — adapting a trained model to a second dataset

Notes

Notebook outputs are committed, so the figures and result tables render on GitHub without running anything. Screening datasets are downloaded through improve_utils; trained weights and scheduler logs are not committed.

Research code from my doctoral work at the University of Nebraska–Lincoln (8 notebooks). Previously hosted at github.com/Ved-Piyush/DeepCDR_TF.

---

Ved Piyush, PhD · Website · Google Scholar · vedpiyush93@gmail.com