OCR: OmniNet-Fusion: A hybrid attention-based CNN-RNN model for multi-omics integration in precision cancer drug response prediction

Abstract

The growing complexity of cancer therapeutics challenges the use of state-of-the-art computational models for drug response prediction. Design and implementation of the OmniNet-Fusion (OCR), a multi-omics deep excavating learning framework for precision medicine. The model uses Convolutional Neural Networks (CNNs) for spatial feature learning and Recurrent Neural Networks (RNNs) for temporal pattern capturing and contains an attention mechanism for focusing on key features among omics layers. Lasso regression and mutual information filter are used for feature selection, and principal component analysis (PCA) enables reduction of the dimension for computing the log p-values. The model was developed based on the CTRPv2 dataset19 which is publicly available. The predictive performance was evaluated based on the experimental results, which were 94.2 % of accuracy, +92.8 % of precision, 91.5 % of recall, and 0.96 of AUC-ROC, indicating superiority over some state-of-the-art baseline methods. Although the OCR model greatly enhances the prediction accuracy and biological interpretability, it also has several issues such as that it requires much more training time because of complex architecture, heavy memory load due to the multi-omics data fusion, and minimal validation in real-time clinical scenarios. Notwithstanding such limitations, OmniNet-Fusion makes a significant contribution towards personalized oncology by providing a scalable and interpretable framework for precision prediction of drug response, while promoting the development of AI-enabled precision medicine.