Several challenges appear in the application of deep learning to genomic data. First, the dimensionality of input can be orders of magnitude greater than the number of samples, forcing the model to be prone to overfitting the training dataset. Second, each input variable's contribution to the prediction is usually difficult to interpret, owing to multiple nonlinear operations. Third, genetic data features sometimes have no innate structure. To alleviate these problems, we propose a modification to Diet Networks by adding element-wise input scaling. The original Diet Networks concept can considerably reduce the number of parameters of the fully-connected layers by taking the transposed data matrix as an input to its auxiliary network. The efficacy of the proposed architecture was evaluated on a binary classification task for lung cancer histology, that is, adenocarcinoma or squamous cell carcinoma, from a somatic mutation profile. The dataset consisted of 950 cases, and 5-fold cross-validation was performed for evaluating the model performance. The model achieved a prediction accuracy of around 80% and showed that our modification markedly stabilized the learning process. Also, latent representations acquired inside the model allowed us to interpret the relationship between somatic mutation sites for the prediction.
Keywords: Diet Networks; deep learning; interpretable neural networks; lung cancer.