Yan Wang Xiaoye Jin Rui Qiu Bo Ma Sheng Zhang Xuyang Song Jinxi He ^*

• General Hospital of Ningxia Medical University, Yinchuan, Ningxia Hui Region, China

Tumor heterogeneity leads to varying patient responses to different drugs, posing a significant challenge in selecting the most suitable treatment. Personalized cancer therapy has emerged as an effective strategy for improving treatment effectiveness and precision. This study aimed to develop a personalized drug recommendation model to improve therapeutic outcomes based on patients’ genomic profiles. Using a content-based filtering algorithm, patient features were characterized by the tumor microenvironment (TME) and drug features by drug fingerprints. The model was initially trained and validated with the GDSC database and further evaluated using the CCLE dataset as an independent validation set. It was then applied to patient data from TCGA, with Best Overall Response (BOR) used as the efficacy measure. Two multilayer perceptron (MLP) models were developed to predict drug sensitivity (IC50) for 542 tumor cell lines across 18 drugs. The model demonstrated high predictive accuracy, with correlation coefficients (R) of 0.914 in the training set and 0.902 in the test set. Sensitivity to cytotoxic drugs, including Docetaxel (R=0.72) and Cisplatin (R=0.71), was particularly well-predicted, though predictions for targeted therapies were less accurate (R<0.3). Validation using CCLE (measuring drug sensitivity via MFI) further supported the model’s robustness (R=0.67). When applied to TCGA data, the model successfully predicted clinical outcomes, with higher predicted scores in the Progressive Disease (PD) group and a significant correlation with 6-month progression-free survival (P=0.007). The area under the curve (AUC) for 6-month PFS was 0.793. Overall, the model shows promise for application in personalized cancer treatment.