Supplementary Figure 8: Correlation patterns of weights mimic BLOSUM62 and Grantham score matrices.
From: Predicting the clinical impact of human mutation with deep neural networks
Correlation patterns of weights from the first three layers of the secondary-structure DL network show correlations between amino acids that are similar to BLOSUM62 and Grantham score matrices. The left heat map shows the correlation of parameter weights from the first convolutional layer following two initial upsampling layers of the secondary-structure DL network between amino acids encoded using a one-hot representation. The middle heat map shows BLOSUM62 scores between pairs of amino acids. The right heat map shows Grantham distance between amino acids. The Pearson correlation between DL weights and BLOSUM62 scores is 0.63 (P = 3.55 × 10–9). The correlation between DL weights and Grantham scores is –0.59 (P = 4.36 × 10–8). The correlation between BLOSUM62 and Grantham scores is –0.72 (P = 8.09 × 10–13).
