jbenbudd/ptm-llama

PTM-LLaMA: Post-Translational Modification Site Prediction

This model, jbenbudd/ptm-llama, is a 7 billion parameter LoRA fine-tune of GreatCaptainNemo/ProLLaMA_Stage_1. It is specifically instruction-tuned to predict post-translational modification (PTM) sites for three types: methylation, phosphorylation, and ubiquitination.

Key Capabilities

• Multi-PTM Prediction: Predicts sites for methylation, phosphorylation, and ubiquitination using a single adapter.
• Sequence-based Analysis: Takes a 21-residue peptide window and a PTM-type instruction to generate a list of modified positions.
• Full-Protein Inference: Employs a three-step sliding window architecture (stride 5) to aggregate per-window predictions into full-protein consensus scores, followed by F1-optimal thresholding.
• Instruction Following: Demonstrates instruction-following capabilities, with cross-instruction ablation showing higher AUC on matched PTM types.

Performance Highlights

Evaluated on a held-out test set, the model achieves notable F1 scores:

• Methylation: 18.28% F1 (AUC 0.613)
• Phosphorylation: 30.95% F1 (AUC 0.653)
• Ubiquitination: 28.11% F1 (AUC 0.765)

Limitations

• Window-local outputs: Predictions are initially window-local, requiring aggregation for full-protein context.
• No structural context: Relies solely on primary sequence, lacking 3D structural information.
• Limited PTM types: Specifically trained and evaluated for only three PTM types.

Good For

• Researchers and developers in bioinformatics needing to predict specific PTM sites.
• Applications requiring automated annotation of methylation, phosphorylation, or ubiquitination in protein sequences.
• Integration into pipelines for protein function prediction or drug target identification where PTMs are critical.