A novel self-attention model based on cosine self-similarity for cancer classification of protein mass spectrometry

Mass spectrometry has become a popular tool for cancer classification. A novel self-attention deep learning model based on cosine self-similarity was proposed to classify cancer by mass spectrometry. First, a primary feature vector is dimensionally reduced by two fully connected layers. Second, the feature vector is transformed into the 2D feature matrix, which can be used to calculate the cosine self-similarity matrix of the self-attention model. Next, three convolutional layers are used to extract the refined feature matrix. Finally, the refined feature matrix is fed into the multi-layer fully-connected network to classify the mass spectra. Experimental results of ovarian and prostate cancer demonstrate that the proposed method outperforms the other methods. •A novel self-attention deep learning model based on cosine self-similarity was proposed.•Primary feature vector can be transformed into 2D feature Matrix.•Cosine self-similarity matrix was involved to construct the self-attention mechanism.•Good classification results of ovarian and prostate cancer can be obtained.