Drugs used in combination therapy has superior treatment efficacy. Preclinical experimental screenings have fueled this progress by identifying and prioritizing new candidate combinations, but they lack predictive power over the transcriptomic data of individualized cancer samples in which exploration of the combination treatment space is intractable. There is an unmet need for computational modeling approaches that can accurately predict personalized synergistic drug combinations by integrating patient-specific molecular data. To overcome this limitation, we developed PAIRWISE, a novel DL framework to survey large preclinical screening dataset for drug combination modeling. Our method showed improved drug synergy prediction compared with benchmarked methods with superior sensitivity/specificity in the prediction of likely synergy. To further investigate safety of drug combination, we developed DDI-GPT, innovated with blending knowledge graph into a sentence as an input in large language model, achieves precision and accuracy in drug-drug interaction prediction.