Discovering molecular features of intrinsically disordered regions by using evolution for contrastive learning
Fig 3
A) The maximum correlation between features in the final convolutional layer and each of the 66 literature-curated features from the trained reverse homology model vs. a randomly initialized model. Features are coloured by their category (top legend). Black trace indicates y = x, while grey traces indicate features more than 2.0x correlated, and less than 0.5x times correlated than the untrained random features. B) Fold enrichment for the set of nearest neighbors using feature representations from the final convolutional layer of the target encoder of our reverse homology model, versus literature-curated feature representations, for 92 GO Slim terms. We show the names of some GO terms in text boxes. C) Area under the receiver operating curve (AUC) for regularized logistic regression classification of mitochondrial targeting signals and Cdc28 targets obtained through 5-fold cross validation. A deep language model (Unirep, gold) performs better than reverse homology (blue) and literature-curated features (green). D) Features with largest coefficients (indicated below each logo) selected by the sparse classifier are consistent with the known amino acid composition biases in mitochondrial targeting signals (left) and short linear motifs in Cdc28 substrates (right).