🧬 ptm-llama — multi-PTM site predictor
Predict post-translational modification sites in a protein sequence. Choose a PTM type and paste an amino-acid sequence; the model runs sliding-window inference, aggregates per-residue consensus votes, and applies the calibrated threshold for that PTM type.
ptm-llama is an instruction-tuned causal protein language model that predicts post-translational modification (PTM) sites from raw amino-acid sequence. A single LoRA adapter was fine-tuned on top of ProLLaMA to jointly handle methylation, phosphorylation, and ubiquitination — the PTM type is selected purely through the natural-language instruction at inference time, demonstrating that a single generative protein LLM can be steered across PTM prediction tasks without task-specific classification heads.
At inference the app runs sliding-window generation (window 21, stride 5), aggregates per-residue consensus scores, and applies a per-PTM-type threshold calibrated on a held-out protein-level split.
Model weights, inference config, calibrated thresholds, and full evaluation (per-PTM metrics, per-residue-type breakdown, cross-instruction ablation) are on Hugging Face: jbenbudd/ptm-llama.
The instruction handed to the model at inference time.
| Amino acid sequence | PTM type |
|---|
Structure predicted on demand via the ESMFold API for sequences ≤ 400 residues. Model weights: jbenbudd/ptm-llama.