InstaDeepAI/ms_ninespecies_benchmark
Viewer β’ Updated β’ 639k β’ 1.07k β’ 3
De novo peptide sequencing from tandem mass spectrometry (MS/MS) using discrete flow matching over Continuous-Time Markov Chains (CTMC).
| Checkpoint Name | Description | Size |
|---|---|---|
frozen_production_model.ckpt |
Canonical Production Model: Pretrained on synthetic peptides from HC-PT (InstaDeepAI/ms_proteometools) and fine-tuned on the Nine-Species benchmark (InstaDeepAI/ms_ninespecies_benchmark, trained on 8 species, evaluated on yeast). | 454 MB |
dfm_massive_kb.ckpt |
MassIVE-KB Pretrained Model: Trained exclusively on ~28M human MS/MS spectra from the MassIVE-KB dataset (MSV000084075) without fine-tuning (supports peptide sequences up to length 50). | 227 MB |
ptm_extended_warmstart.ckpt |
PTM Extended Model: Pretrained base model with extended post-translational modification (PTM) vocabulary. | 476 MB |
import torch
from huggingface_hub import hf_hub_download
from train.factory import build_models
from train.io import load_checkpoint, load_models_from_checkpoint
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Download and load checkpoint (e.g. frozen_production_model.ckpt or dfm_massive_kb.ckpt)
ckpt_path = hf_hub_download(repo_id="joelinator/dflow-novo-model", filename="frozen_production_model.ckpt")
ckpt = load_checkpoint(ckpt_path, map_location=device)
vocab = ckpt["vocabulary"]
# Build & load models
encoder, len_pred, decoder, guidance = build_models(vocab, device)
load_models_from_checkpoint(ckpt, encoder, len_pred, decoder, guidance, use_ema=True)
encoder.eval(); len_pred.eval(); decoder.eval(); guidance.eval()
Try the live Gradio interface at: https://huggingface.co/spaces/joelinator/dflow-novo
@mastersthesis{gedeon2026dflownovo,
title={Discrete Flow Matching for De Novo Peptide Sequencing in Mass Spectrometry},
author={G{\'e}d{\'e}on, Jo{\"e}l},
school={African Institute for Mathematical Sciences (AIMS South Africa)},
year={2026}
}