Multi-function Prediction of Unknown Protein Sequences Using Multilabel Classifiers and Augmented Sequence Features

Pretrained language models and weight redistribution achieve precisekcatprediction

“…Then, these embedded vectors generate a per-molecular representation for a substrate. Furthermore, the concatenated representation vector of protein and substrate are imputed into the explainable extra tree model, which has been clarified to be useful on various protein or peptide function prediction tasks [19][20]. We also found that a solely concatenated representation vector cannot discriminate high or low k cat values well by projection of t-distributed stochastic neighbour embedding (t-SNE) [21], further demonstrating the necessity of the machine learning model (Supplementary Fig.…”

“…DM means that the weight of samples with k cat values higher than 5 (logarithm value) would be enhanced. We compared several parameters, including the weight multipliers (5,10,20,50) and whether they were normalized. This resulted in eight optimized model combinations.…”