Abstract
Background and Objective: This paper aimed to differentiate primary cancer types from primary tumor samples on the basis of somatic point mutations (SPMs). Primary cancer site identification is necessary to perform site-specific and potentially targeted treatment. Current methods such as histopathology and lab tests cannot accurately determine cancer origin, which results in empirical patient treatment and poor survival rates. The availability of large deoxyribonucleic acid sequencing datasets has allowed scientists to examine the ability of somatic mutations to classify primary cancer sites. These datasets are highly sparse since most genes will not be mutated, have a low signal-to-noise ratio, and are often imbalanced since rare cancers have fewer samples. Methods: To overcome these limitations a sparse-input neural network (SPINN) is suggested that projects the input data in a lower-dimensional space, where the more informative genes are used for learning. To train and evaluate SPINN, an extensive dataset for SPM was collected from the cancer genome atlas containing 7624 samples spanning 32 cancer types. Different sampling strategies were performed to balance the dataset. SPINN was further validated on an independent ICGC dataset that contained 226 samples spanning four cancer types. Results and Conclusions: SPINN consistently outperformed classification algorithms such as extreme gradient boosting, deep neural networks, and support vector machines, achieving an accuracy up to 73% on independent testing data. Certain primary cancer types/subtypes (e.g., lung, brain, colon, esophagus, skin, and thyroid) were classified with an F-score > 0.80.