A semi-supervised learning approach for pan-cancer somatic genomic variant classification

Cancer arises from the accumulation of particular somatic genomic variants known as drivers. New sequencing technologies allow the identification of hundreds of variants in a tumor sample. These variations should be classified as driver or passenger (i.e. benign), but functional studies could be time and cost demanding. Therefore, in the bioinformatics field, machine learning methods are widely applied to distinguish drivers from passengers. Recent projects, such as the AACR GENIE, provide an unprecedented amount of cancer data that could be exploited for the training process of machine learning algorithms. However, the majority of these variants are not yet classified. The development and application of approaches able to assimilate unlabeled data are needed in order to fully benefit from the available omics-resources. We collected and annotated a dataset of known 976 driver and over 84,000 passengers from different databases and we investigated whether unclassified variants from GENIE could be employed in the classification process. We characterized each variant by 94 features from multiple omics resources. We therefore trained different autoencoder architectures with more than 80000 GENIE variants. Autoencoder is a type of neural network able to learn a new features representation of the input data in an unsupervised manner. The trained autoencoders are then used to obtain new representations of the labeled dataset, with a reduced number of meta-features with the aim to reduce redundancy and extract the relevant information. The new representations are in turn exploited to train and test different machine learning techniques, such as Random Forest, Support Vector Machine, Ridge Logistic Regression, One Class SVM. Final results, however, does not show a significant increase in classification ability when meta-features are used.