A semi-supervised weighted SPCA and convolution KANs based Drug response prediction model

Rui Miao ^1* Bing-Jie Zhong ¹ Xin-Yue Mei ² Xin Dong ² Yang-Dong Ou ³ Yong Liang ⁴ Hao-Yang Yu ¹ Ying Wang ¹ Zi-Han Dong ¹

• ¹ Zunyi Medical University, Zunyi, China
• ² Macau university of science and technology, Macau, Macao, SAR China
• ³ Guangdong Medical University, Zhanjiang, Guangdong, China
• ⁴ Peng Cheng Laboratory, Shenzhen, Guangdong Province, China

Motivation: How to predict the response of cell lines to characteristic drugs based on multi-omics gene information has become the core problem of precision oncology. At present, drug response prediction with multi-omics gene data mainly faces the following three challenges: First, how to design a gene probe feature extraction model with biological interpretation and high performance. Second, how to design multi-omics weighting modules for reasonable fusion of genetic data of different lengths and noise conditions. Third, how to design deep learning models in the case of small samples to avoid the risk of possible overfitting.We propose an innovative drug response prediction model (NMDP). Firstly, NMDP model proposes an interpretable semi-supervised weighted SPCA module to solve the feature extraction problem of multi-omics gene data. Next, we construct a multi-omics data fusion framework based on sample similarity network, bimodal test and variance information, which solves the data fusion problem and enables the NMDP model to focus on more important genomic data. Finally, we combined onedimensional convolution and KANs (Kolmogorov-Arnold Networks) method to predict drug response. We conduct 5 sets of real data experiments and compare 7 advanced drug response prediction methods which show that NMDP has the best performance, its sensitivity and specificity can reach 0.92 and 0.93, due to the comparison model 11%-57%. Bioenrichment experiments strongly proved that the NMDP model has good biological interpretation and can select potential drug activity prediction targets.