Pretraining on Phosphosite Sequences with MLM Objective on ESM-1b Architecture
This repository provides a pretrained ESM-1b architecture on phosphosite sequences, where the weights are initialized from scratch and trained using the Masked Language Modeling (MLM) objective. The model was trained on long phosphosite-containing peptide sequences derived from PhosphoSitePlus.
Developed by:
Zeynep Işık (MSc, Sabanci University)
Training Details
Architecture: ESM-1b (trained from scratch)
Pretraining Objective: Masked Language Modeling (MLM)
Dataset: Unlabeled phosphosites from PhosphoSitePlus
Total Samples: 352,453 (10% seperated for validation)
Sequence Length: ≤ 128 residues
Batch Size: 64
Optimizer: AdamW
Learning Rate: default
Training Duration: 1.5 day
Pretraining Performance
Perplexity at Start: 13.51
Perplexity at End: 2.20
A significant decrease in perplexity indicates that the model has effectively learned meaningful representations of phosphosite-related sequences.
Potential Usecases
This pretrained model can be used for downstream tasks requiring phosphosite knowledge, such as:
✅ Binary classification of phosphosites
✅ Kinase-specific phosphorylation site prediction
✅ Protein-protein interaction prediction involving phosphosites
