FIGURE 1.

The AAIndex database of amino acid physicochemical properties can be split into eight semantically-consistent groups. (A) Combining doc2vec strategies with unsupervised learning algorithms, we proposed a methodology to generate groups that preserve semantic consistency within the partition. Applying an RBF kernel PCA on the whole dataset, we observe that the groups are linearly separable in the PCA1/PCA2 space, as their convex hulls are disjoint. (B,C) Combining our encoders with FFT improves model performance and helps reducing overfitting in several predictive tasks. Here, boxplots summarize the distribution of performances reached in each experiment across the 1,000 independent realizations of the 80/20 split of the input dataset for the task. Central circles represent medians, bars the interquartile range, and whiskers the 95% CI. Complementary analyses of model performance, including other metrics (such as recall, F-Score, and area under the receiver operating curves AUC), are presented in Supplementary Section S3 and summarized in Supplementary Tables S3–S6.