PhosBoost

PhosBoost predicts protein phosphorylation sites using protein language models and gradient-boosting trees to improve identification of serine, threonine, and tyrosine phosphosites, with emphasis on plant proteins.

Key Features:

• Model architecture: Combines protein language models and gradient-boosting trees for phosphosite prediction.
• Training data: Trained on phosphorylation annotations from the qPTMplants database.
• Performance metrics: Demonstrated recall of 0.78 for serine and threonine sites versus 0.56 and 0.14 for comparative methods, recall of 0.6 for tyrosine sites where other tools failed, and a reported precision-recall score of 0.54 with more confident probability scores.
• Sequence alignment enhancement: Incorporates a sequence-based pairwise alignment step that increases the inferred number of positive phosphosites across classifiers.

Scientific Applications:

• Research in plant biology: Enables genome-wide prediction of phosphorylation sites to support studies of plant signaling pathways and protein interactions given limited experimental data.
• Cross-species transferability: Models show evidence of transferability across species, extending applicability beyond the original training datasets.

Methodology:

Uses protein language models and gradient-boosting trees trained on qPTMplants, includes a sequence-based pairwise alignment step, and was evaluated against PhosphoLingo and DeepPhos with reported recall and precision-recall metrics.