InstaNovo-P: A De Novo Peptide Sequencing Model for Phosphoproteomics

InstaNovo-P is a phosphorylation-specific version of the transformer-based InstaNovo model, fine-tuned on extensive phosphoproteomics datasets. It significantly surpasses existing methods in phosphorylated peptide detection and phosphorylation site localization accuracy across multiple datasets.

Paper: InstaNovo-P: A de novo peptide sequencing model for phosphoproteomics (bioRxiv preprint)

Important -- Codebase Lineage: This repository contains a self-contained fork of InstaNovo v0.1.6 with substantial modifications for phosphoproteomics fine-tuning. It is not compatible with newer InstaNovo releases (>=1.0).

For inference only, use the latest instanovo package (>=1.1.2) with the released InstaNovo-P checkpoint. See the Quick Start -- Inference section below.

This repository provides the training and fine-tuning code to reproduce the results presented in the paper.

Table of Contents

• Quick Start -- Inference
• Datasets
• Codebase Lineage and Modified Files
• Reproducing the Fine-Tuning
• Evaluation
• Hyperparameters
• Citation
• License

Quick Start -- Inference

For running inference with InstaNovo-P, we recommend using the latest instanovo package which provides a streamlined interface:

pip install instanovo>=1.1.2

The InstaNovo-P checkpoint is available at:

• instanovo-phospho-v1.0.0.ckpt

For a step-by-step inference tutorial, see the official notebook:

• InstaNovo-P.ipynb

The notebook demonstrates:

• Loading the InstaNovo-P checkpoint
• Loading phosphoproteomics data from HuggingFace
• Running inference with greedy, beam search, and knapsack beam search decoding
• Evaluating prediction confidence and computing metrics

Datasets

Training Data

The InstaNovo-P training dataset is available on HuggingFace:

• InstaDeepAI/InstaNovo-P

This dataset comprises 2.57 million phosphorylated peptide-spectrum matches (PSMs) from 29 PRIDE projects, reprocessed with IonBot (filtered at 0.80 localization probability) and split using GraphPart homology partitioning into train/validation/test sets.

Evaluation Datasets

• 21PTM Validation: https://huggingface.co/datasets/InstaDeepAI/PXD009449
• FGFR2 Validation: https://huggingface.co/datasets/InstaDeepAI/PXD062859
• Astral Validation: https://huggingface.co/datasets/InstaDeepAI/PXD055983

Codebase Lineage and Modified Files

This repository is based on InstaNovo v0.1.6 with extensive modifications for phosphoproteomics fine-tuning. Below is a summary of every file and its changes relative to the public instanovo==0.1.6 release.

Modified Files

File	Diff Size	Description
instanovo/transformer/train.py	610 lines	Core training loop rewritten for phospho fine-tuning. Adds gradual unfreezing via FinetuningScheduler, slanted triangular learning rate scheduling (STLRScheduler), PTM-specific metric logging (phospho precision/recall), checkpoint key remapping for the extended residue vocabulary, and Polars IPC / HuggingFace data loading.
instanovo/transformer/predict.py	555 lines	Prediction pipeline adapted for phospho residues. Adds PTM residue remapping between internal notation (S(+79.97)) and ProForma (S[UNIMOD:21]), evaluation on phospho-specific metrics, and flexible checkpoint loading with vocabulary extension support.
instanovo/transformer/model.py	368 lines	Model architecture extended with optional LoRA adapter support for parameter-efficient fine-tuning, checkpoint loading logic for vocabulary-extended models (mapping new phospho tokens into the embedding/output layers), and decode helper methods.
instanovo/utils/metrics.py	284 lines	Evaluation metrics extended with PTM-specific precision and recall (measuring correct localization of phosphorylation and oxidation sites independently of overall peptide accuracy).
instanovo/utils/residues.py	197 lines	Residue vocabulary extended with phosphorylated residues (S(+79.97), T(+79.97), Y(+79.97)) and oxidized methionine (M(+15.99)). Adds bidirectional mapping between internal and ProForma/UNIMOD notation.
instanovo/transformer/dataset.py	79 lines	Dataset class extended with string-return mode for sequence labels, flexible column name mapping, and Polars IPC format support alongside the original MGF-based loading.
instanovo/utils/convert_to_ipc.py	52 lines	Data conversion script updated to handle the phospho dataset schema.
instanovo/inference/beam_search.py	25 lines	Minor adjustments to beam search decoder.
instanovo/constants.py	14 lines	Added SpecialTokens enum (PAD, EOS, SOS).
instanovo/inference/knapsack_beam_search.py	4 lines	Trivial changes.

Unchanged Files

These files are identical to their instanovo==0.1.6 counterparts:

• instanovo/transformer/layers.py -- Transformer encoder/decoder layer definitions
• instanovo/inference/knapsack.py -- Knapsack algorithm for constrained beam search
• instanovo/inference/interfaces.py -- Abstract interfaces for decoders

New Files (not in v0.1.6)

File	Description
instanovo/utils/schedulers.py	Implements STLRScheduler and STLRSchedulerEpochBased for slanted triangular learning rate scheduling used during gradual unfreezing.
instanovo/utils/minLoRA.py	Lightweight LoRA (Low-Rank Adaptation) implementation for parameter-efficient fine-tuning of transformer layers.

Reproducing the Fine-Tuning

1. Clone the Repository

git clone https://github.com/<OWNER>/instanovo-p.git
cd instanovo-p

2. Install Dependencies

This project uses uv for dependency management with Python 3.10 and pinned dependency versions for reproducibility.

uv venv --python 3.10
source .venv/bin/activate
uv pip install -e .

3. Download the Base InstaNovo Checkpoint

The fine-tuning starts from the base InstaNovo model (v0.1.4):

wget https://github.com/instadeepai/InstaNovo/releases/download/0.1.4/instanovo.pt -P checkpoints/

This downloads the 361 MB checkpoint to ./checkpoints/instanovo.pt.

4. Download and Prepare Training Data

Download the training data from HuggingFace:

from datasets import load_dataset

dataset = load_dataset("InstaDeepAI/InstaNovo-P")

Convert the data to Polars IPC format for training:

python -m instanovo.utils.convert_to_ipc --input <path_to_data> --output data/

5. Run Training

python -m instanovo.transformer.train \
    --train_path data/train.ipc \
    --valid_path data/valid.ipc \
    --config instanovo_finetune_phospho.yaml \
    --n_workers 8

The training configuration is defined in configs/instanovo/instanovo_finetune_phospho.yaml and uses the gradual unfreezing schedule from configs/finetune_scheduler/finetune_schedule_gu_decoder-encoder-v2.yaml.

The fine-tuning proceeds through 10 epochs with encoder-first gradual unfreezing:

1. Epochs 0-1: Head and embedding layers only
2. Epochs 1-5: Decoder layers progressively unfrozen (top to bottom)
3. Epochs 5-10: Encoder layers progressively unfrozen (top to bottom)

Evaluation

Run prediction on a test set:

python -m instanovo.transformer.predict \
    <test_data.ipc> \
    <checkpoint_path> \
    --output-path predictions.csv \
    --n-beams 5 \
    --batch-size 128

Evaluation metrics (amino acid precision/recall, peptide precision/recall, PTM-specific precision/recall) are computed during validation and logged to TensorBoard. The predictions CSV can be used for further downstream analysis.

Hyperparameters

The key hyperparameters used for fine-tuning InstaNovo-P:

Parameter	Value
Base model	InstaNovo v0.1.4 (768-dim, 9 layers, 16 heads)
Optimizer	Adam
Learning rate scheduler	STLR (Slanted Triangular LR), epoch-based
Max learning rate	2e-4 (head), 1e-5 to 2e-6 (deeper layers)
LR ratio	1/32
Batch size	64
Epochs	10
Weight decay	1e-6
Dropout	0.1
Gradient clipping	10.0
Unfreezing strategy	Gradual (decoder-first, then encoder)
Validation beams	2
Training samples	~1.89M
Validation subset	2% (~4,437 samples)

Citation

If you use InstaNovo-P in your research, please cite:

@article {Lauridsen2025.05.14.654049,
  author = {Lauridsen, Jesper and Ramasamy, Pathmanaban and Catzel, Rachel and Canbay, Vahap and Mabona, Amandla and Eloff, Kevin and Fullwood, Paul and Ferguson, Jennifer and Kirketerp-M{\o}ller, Annekatrine and Goldschmidt, Ida Sofie and Claeys, Tine and van Puyenbroeck, Sam and Lopez Carranza, Nicolas and Schoof, Erwin M. and Martens, Lennart and Van Goey, Jeroen and Francavilla, Chiara and Jenkins, Timothy Patrick and Kalogeropoulos, Konstantinos},
  title = {InstaNovo-P: A de novo peptide sequencing model for phosphoproteomics},
  elocation-id = {2025.05.14.654049},
  year = {2025},
  doi = {10.1101/2025.05.14.654049},
  publisher = {Cold Spring Harbor Laboratory},
  URL = {https://www.biorxiv.org/content/early/2025/05/18/2025.05.14.654049},
  eprint = {https://www.biorxiv.org/content/early/2025/05/18/2025.05.14.654049.full.pdf},
  journal = {bioRxiv}
}

Note: This citation refers to the bioRxiv preprint. It will be updated once the peer-reviewed publication is available.

License

This project is licensed under the Apache License 2.0 -- see the LICENSE file for details.