An In-Memory Cognitive-Based Hyperdimensional Approach to Accurately Classify DNA-Methylation Data of Cancer.

With Next Generation DNA Sequencing techniques (NGS) we are witnessing a high growth of genomic data. In this work, we focus on the NGS DNA methylation experiment, whose aim is to shed light on the biological process that controls the functioning of the genome and whose modifications are deeply investigated in cancer studies for biomarker discovery. Because of the abundance of DNA methylation public data and of its high dimension in terms of features, new and efficient classification techniques are highly demanded. Therefore, we propose an energy efficient in-memory cognitive-based hyperdimensional approach for classification of DNA methylation data of cancer. This approach is based on the brain-inspired Hyperdimensional (HD) computing by adopting hypervectors and not single numerical values. This makes it capable of recognizing complex patterns with a great robustness against mistakes even with noisy data, as well as the human brain can do. We perform our experimentation on three cancer datasets (breast, kidney, and thyroid carcinomas) extracted from the Genomic Data Commons portal, the main repository of tumoral genomic and clinical data, obtaining very promising results in terms of accuracy (i.e., breast 97.7%, kidney 98.43%, thyroid 100%, respectively) and low computational time. For proving the validity of our approach, we compare it to another state-of-the-art classification algorithm for DNA methylation data. Finally, processed data and software are freely released at https://github.com/fabio-cumbo/HD-Classifier for aiding field experts in the detection and diagnosis of cancer.