Abstract
In clinical practice, the development of personalized treatment strategies for cervical cancer is hindered by the limited accuracy of drug response prediction, partly due to missing modalities in multi-omics data. We present MKDR, a deep learning framework that integrates variational autoencoder-based modality completion with knowledge distillation to transfer information from complete omics data to incomplete samples. MKDR-Student achieves state-of-the-art performance On cervical cancer cell lines, with an MSE of 0.0034 (34% lower than Xgboost), R² of 0.8126, and MAE of 0.0431, while maintaining high Spearman (0.8647) and Pearson (0.9033) correlations. Data ablation experiments highlight the contributions of knowledge distillation and modality completion: removing the teacher increases MSE by 23%, and VAE reduces error by 15% with 40% missingness. Interpretability analysis shows balanced feature contributions from gene expression (38%), copy number variation (30%), and mutation data (32%), indicating effective multi-omics learning and integration by the student model. Under limited-input conditions, MKDR's accuracy drops less than 5%, supporting its robustness and potential for clinical application.