The rapid growth of precision oncology demands computational tools capable of uncovering meaningful patterns between chemical compounds and cancer-specific drug responses. Leveraging high-throughput biological data and cheminformatics, we developed a predictive modeling framework using GI50 data from the NCI-60 cancer cell line panel to classify compound activity and to characterize structure function relationships. Compounds are encoded using SMILES notation and processed through RDKit to generate Morgan fingerprints, which represent molecular substructures as binary vectors. Datasets from the NCI-60 panel were selected to train and test Support Vector Machine (SVM) classifiers, distinguishing active from inactive compounds based on GI50 concentration thresholds. To improve both model performance and interpretability, we integrated MISTIC (Model Informed Selection Through Importance and Contribution), a custom feature selection method designed to identify the most informative fingerprint bits. Additionally, Tanimoto similarity scores are used to define the SVM kernel, enhancing the model's ability to capture structural relationships between compounds. Our approach demonstrates strong predictive accuracy and revealed key molecular features linked to drug efficacy. These results highlight the power of combining cheminformatics with machine learning to support personalized treatment strategies in precision oncology. Future work will extend this framework to additional cancer types and incorporate genomic biomarkers for deeper clinical insight.
