Deep learning-based metabolomics data study of prostate cancer

Abstract Background: As a heterogeneous disease, prostate cancer (PCa) presents diverse clinical and biological features that pose challenges for early diagnosis and treatment. Metabolomics can provide new methods for the early diagnosis, treatment, and prognosis of prostate cancer. However, metabolomics data are characterized by high throughput, sparsity, high dimensionality, and small samples, which poses great challenges for classification. Despite the wide range of applications of deep learning methods, metabolomics research based on deep learning has not yet been widely conducted. Results: In this study, we propose a hybrid model, namely, TCNet, based on transformer and convolutional neural networks for prostate cancer metabolomics data classification. We introduce a 1D convolution for the inputs of the dot product attention, which enables the interaction of local-global information. A gating mechanism is used so that the model can dynamically adjust the attention weights. The features extracted by the multi-head attention are extracted at a more advanced level via 1D convolution. A residual network is introduced in the 1D convolution to alleviate the gradient vanishing problem. A five-fold cross-validation was used to complete the classification experiment, our experimental study shows that the TCNet model based on a transformer with a convolutional neural network can obtain better classification results than seven other machine learning comparison algorithms. Conclusions: We propose a novel hybrid model for classifying prostate cancer metabolomics data and demonstrate that our approach outperforms other methods based on deep learning. Our results provide new perspectives on prostate cancer early diagnosis.