raowusu15/west-african-cardiotoxicity-screening

# West African Medicinal-Plant Cardiotoxicity Screening An exploratory computational project evaluating potential off-target cardiotoxicity of West African medicinal-plant constituents across three cardiac ion channels: **hERG (KCNH2), Nav1.5 (SCN5A), and Cav1.2 (CACNA1C)**. The project was inspired by the three-channel cardiotoxicity framework and public dataset reported by **Arab et al.**, but adapts the workflow to a West African medicinal-plant compound collection rather than developing a general-purpose CToxPred replacement. > **Portfolio status:** research-in-progress. Outputs are computational predictions for > prioritization and hypothesis generation, not clinical safety labels. ## Research question Can conventional Morgan-fingerprint machine-learning models identify West African phytochemicals that may warrant closer assessment for hERG, Nav1.5, or Cav1.2 blockade? ## Workflow 1. Load target-specific binary cardiotoxicity training data. 2. Encode SMILES as 1024-bit Morgan fingerprints (radius = 2). 3. Evaluate KNN, SVC and MLP classifiers using a stratified 70/30 split. 4. Retrain selected models on the full target dataset. 5. Screen 1,183 West African compound records. 6. Map predictions to a 54-herb subset containing 383 constituent records. 7. Compare model agreement and flag high-priority candidates for follow-up. ## Model architecture and evaluation - **KNN:** k in {3, 5, 7, 9}, 5-fold ROC-AUC grid search. - **SVC:** probability-enabled RBF SVC with grid search over C/gamma. - **MLP:** 1024 -> 512 -> 256 -> 1, ReLU hidden layers, dropout 0.2, sigmoid output, Adam (0.001), 20 epochs, batch size 64. - Metrics: sensitivity, specificity, accuracy, balanced accuracy, MCC, AUC-ROC and F1. ## Stored evaluation results | Target | Model | AUC-ROC | MCC | Accuracy | F1 | |---|---:|---:|---:|---:|---:| | hERG | KNN | 0.861 | 0.570 | 0.786 | 0.767 | | hERG | MLP | **0.881** | **0.605** | **0.803** | **0.793** | | Nav1.5 | KNN | 0.874 | 0.599 | 0.800 | 0.78 …