Overview

This project develops an explainable multimodal model that predicts whether a small-molecule compound will inhibit a paediatric cancer cell line (IC₅₀ ≤ 0.368 µM).

Test-set metric	Value
AUROC	0.981
Precision	0.973
Recall	0.985

Data sources

• CCLE 22Q2 — 735 high-variance gene-expression features

• GDSC2 — 108,696 drug–cell-line IC₅₀ measurements

• SMILES fingerprints — 2,048-bit Morgan fingerprints

Model at a glance

Two dense branches (genes: 735 → 128, chemistry: 2,048 → 128) feed a multi-head cross-modal attention layer; the fused representation passes through three fully connected layers and a final sigmoid unit to predict sensitivity.

Key links

• GitHub repository — https://github.com/htootayzaaung/Explainable-AI-Driven-Adverse-Drug-Reactions-Prediction-Toward-Pediatric-Drug-Discovery-Development

• Google Drive — full project folder with datasets, notebooks, trained model, and logs — https://drive.google.com/drive/folders/1PXvxMIHJs_etK4K1nf5Hm4fkfUrDreaO