"Re-evaluating pIC50 Predictive Limits: Why R² = 0.74 is the True Gold Standard for Honest Bioactivity Modeling"

This repository contains the complete dataset, Out-of-Fold (OOF) predictions, and methodology for the pIC50 potency prediction model developed by the ActarusLab Research Team (March 2026). The study addresses the critical issue of over-optimistic performance metrics in QSAR (Quantitative Structure-Activity Relationship) modeling. While previous benchmarks claimed an R2 of 0.86 on similar ChEMBL datasets, we demonstrate through a rigorous validation pipeline that a more realistic and honest baseline is R2=0.738. Key Technical Features: Architecture: A hybrid ensemble combining Graph Attention Networks (GAT) for molecular topology with Gradient Boosting machines (XGBoost, LightGBM, CatBoost) for tabular molecular features. Validation Protocol: 5-fold cross-validation repeated with 3 different random seeds (15 total evaluations). All metrics are calculated strictly Out-of-Fold (OOF) to prevent data leakage. Target: pIC50 values extracted from ChEMBL v33, clipped in the interval [−0.5,4.0] to reduce experimental noise impact. Performance: R2=0.738, RMSE=0.597, MAE=0.472. Contents of this upload: predictions_multiseed_r2_074.csv: The full set of experimental vs. predicted values. pIC50_ensemble_notebook.ipynb: The complete Python pipeline (optimized for T4 GPUs). paper_pIC50_ensemble_ITA.docx: The full technical report.