CandyCrunch v2 Benchmark (rlacombe/candycrunch-v2-benchmark)

Leak-free, LC-MS run-grouped benchmark for single-spectrum de novo glycan structure prediction from tandem mass spectrometry (MS/MS), accompanying "GlycoGen: Glycan Structure Prediction with Crystallizing Flows" (NeurIPS 2026; code repository).

Derived deterministically from the 503,446-spectrum CandyCrunch v2 corpus curated by Urban et al. (Nature Methods, 2024) and archived at Zenodo 10.5281/zenodo.15741423 (Abtheen et al., 2025) under CC BY 4.0.

---

Overview and motivation

In liquid chromatography-tandem mass spectrometry (LC-MS/MS) glycomics, a single glycan eluting off a porous graphitized carbon (PGC) or C18 column generates 5 to 50 repeated MS/MS scans within one analytical run across adjacent retention times, multiple charge states, adduct forms, and chromatographically separated $\alpha/\beta$ reducing-end anomers.

Under a uniform random row-level train/test split, sibling scans of the same chromatographic peak land in both training and test sets. Models evaluated on random spectrum splits can therefore exploit memorized retention times, run-specific ionization profiles, and biological sample matrices rather than generalizing to unseen runs—inflating top-1 accuracy by 2 to 5 percentage points.

Following Urban et al. (2024), CandyCrunch v2 Benchmark partitions the corpus strictly at the LC-MS source filename level via GroupShuffleSplit(test_size=0.15, n_splits=1, random_state=42), guaranteeing zero shared LC-MS runs between {train, val} and test. Every row is shipped in two synchronized formats sharing identical row order (row_manifest.csv):

1. Continuous dual-channel numerical spectra (X_*.pkl, y_*.pkl) — 2,048-bin intensity + sub-bin m/z remainder vectors (39.714–3000.0 Da) paired with full IUPAC-condensed glycan strings, used by GlycoGen-DFM, GlycoGen-MDLM, and GlycoGen-AR.
2. Tokenized MS sentences and antenna paths (sentences_benchmark.pkl, vocab_glycobart.txt) — quantized m/z, retention-time, and metadata token sequences paired with antenna-decomposed targets, enabling controlled from-scratch retraining of GlycoBART (Abtheen et al., 2025) on the exact same train/val/test rows.

---

Split statistics

Starting from 503,446 raw spectra in dataset.xlsx / dataset.parquet:

1. Run-quality filter (RT.fillna(15) > 2, non-null filename, non-empty L1-normalizable peak_d) retains 499,210 rows.
2. Antenna-mapping filter (excludes curly-brace {...} glycans with unlocalized substituents via glycans_antennae.csv so sequence-to-antenna baselines train on identical rows) retains 501,987 rows.
3. Intersection: 497,751 usable spectra (5,695 rows = 1.13% excluded overall) spanning 3,572 unique IUPAC-condensed glycan strings across O-glycans, N-glycans, free oligosaccharides, and glycosphingolipids.

Split	Spectra	Unique LC-MS filename runs	Shared filenames with test	Assignment rule
train	378,880	2,455 (across train + val)	0	90% of non-test pool (np.random.default_rng(seed=42))
val	42,097	subset of the 2,455 non-test runs	0	10% of non-test pool (np.random.default_rng(seed=42))
test	76,774	434 disjoint runs	—	GroupShuffleSplit(test_size=0.15, random_state=42) by filename
Total	497,751	2,889	0	Deterministic, byte-reproducible

Within the 76,774 held-out test spectra, 76,203 match a training string verbatim, 66 represent training chemical graphs serialized under an unseen IUPAC branch traversal order, and 505 represent glycan graph topologies absent from train.

---

Repository files and data schema

File	Format	Description
X_train.pkl, X_val.pkl, X_test.pkl	Python list of 9-tuples (pickle)	Numerical spectrum features aligned to row_manifest.csv (378,880 / 42,097 / 76,774 rows).
y_train.pkl, y_val.pkl, y_test.pkl	Python list[str] (pickle)	Ground-truth IUPAC-condensed glycan strings in identical row order (e.g. "Neu5Ac(a2-6)Gal(b1-4)GlcNAc").
row_manifest.csv	CSV (497,751 rows)	Per-row provenance linking all formats back to the raw Zenodo table: columns orig_idx,filename,glycan,split.
sentences_benchmark.pkl	Python dict (pickle)	Tokenized MS sentences and antenna targets for GlycoBART reproduction (schema below).
vocab_glycobart.txt	Plain text (10,071 lines)	One token per line built strictly from train (mz* bins, rt* bins, metadata literals, and antenna tokens).
spectrum_metadata.npz	NumPy archive (.npz)	Per-test-row arrays (test_reducing_mass, test_mode, test_modification) used only by optional precursor-mass post-processing diagnostics.
test_peak_d.pkl	Python list[dict] (pickle)	Raw peak dictionaries {mz: intensity} for test rows, used only for raw-preprocessing latency microbenchmarks and CandyCrumbs fragment-reranking ablations.

Schema of X_{train,val,test}.pkl

Each element x = X_split[i] is a 9-tuple following CandyCrunch's numerical format:

Index	Field	Type / Shape	Encoding
x[0]	binned_intensities	np.ndarray, float32[2048]	L1-normalized fragment intensities summed into 2,048 linear bins across [39.714, 3000.0] Da (np.linspace(39.714, 3000.0, 2048)).
x[1]	mz_remainder	np.ndarray, float32[2048]	Sub-bin mass offset mz - frame_edge (Da) of the most intense peak in each bin (0.0 for empty bins), preserving exact peak mass within the ~1.45 Da bin width.
x[2]	reducing_mass	float	Observed precursor m/z value.
x[3]	glycan_type	int	0 = O-glycan, 1 = N-glycan, 2 = free/other reducing end, 3 = fallback.
x[4]	RT_normalized	float	Per-file normalized retention time RT / max(max_file_RT, 30.0).
x[5]	mode	int	Ionization polarity: 0 = negative, 1 = positive, 2 = unknown/other.
x[6]	lc	int	Liquid chromatography column: 0 = PGC, 1 = C18, 2 = unknown/other.
x[7]	modification	int	Sample derivatization: 0 = reduced, 1 = permethylated, 2 = underivatized/other.
x[8]	trap	int	Mass analyzer family: 0 = linear ion trap, 1 = orbitrap, 2 = Bruker amazon, 3 = other/unknown.

Note: GlycoGen-DFM, GlycoGen-MDLM, and GlycoGen-AR consume only x[0], x[1], and x[2] (spectrum intensities, sub-bin m/z remainders, and precursor m/z), without feeding glycan class, retention time, or instrument metadata into the model. Fields x[3..8] are retained for per-instrument evaluation slicing (e.g., x[8] mass-analyzer transfer) and compatibility with CandyCrunch pipelines.

Schema of sentences_benchmark.pkl

A single Python dict with keys:

• train_ms, val_ms, test_ms (list[str]): space-separated CABSEL input sentences (LC_type mode ionization modification trap fragmentation glycan_type rt{bin} mz{prec_bin} mz{peak_1} ...).
• train_glycan, val_glycan, test_glycan (list[str]): space-separated antenna-decomposed target token strings.
• train_glycan_iupac, val_glycan_iupac, test_glycan_iupac (list[str]): full IUPAC-condensed target strings (identical to y_{train,val,test}.pkl).
• vocab_list (list[str]): sorted 10,071-token training vocabulary.
• split_metadata (dict): split seeds, fractions, row counts, and filenames_in_train_and_test = 0 verification record.

---

Quickstart: loading splits in Python

import pickle
from pathlib import Path
import pandas as pd
from huggingface_hub import hf_hub_download

REPO_ID = "rlacombe/candycrunch-v2-benchmark"

def load_split(split: str = "test"):
    x_path = hf_hub_download(repo_id=REPO_ID, filename=f"X_{split}.pkl", repo_type="dataset")
    y_path = hf_hub_download(repo_id=REPO_ID, filename=f"y_{split}.pkl", repo_type="dataset")
    X = pickle.loads(Path(x_path).read_bytes())
    y = pickle.loads(Path(y_path).read_bytes())
    return X, y

X_test, y_test = load_split("test")
print(f"Loaded {len(X_test):,} test spectra")  # 76,774

intensities, mz_remainder, precursor_mz = X_test[0][:3]
print(intensities.shape, mz_remainder.shape, precursor_mz, y_test[0])
# (2048,) (2048,) 1136.42 '...'

# Optional: inspect LC-MS filename provenance
manifest_path = hf_hub_download(repo_id=REPO_ID, filename="row_manifest.csv", repo_type="dataset")
manifest = pd.read_csv(manifest_path)
print(manifest["split"].value_counts().to_dict())
# {'train': 378880, 'test': 76774, 'val': 42097}

---

Evaluation metric

Because branched glycans admit multiple valid IUPAC-condensed string serializations depending on bracket traversal order, string equality undercounts structurally exact predictions. All models on this benchmark are scored per spectrum via glycowork chemical graph isomorphism (Thomès et al., 2021), without multi-scan consensus voting or reference-database snap-back:

from glycowork.motif.graph import compare_glycans

def struct_equal(pred_iupac: str, true_iupac: str) -> bool:
    if not pred_iupac or not true_iupac:
        return False
    if pred_iupac == true_iupac:
        return True
    try:
        return bool(compare_glycans(pred_iupac, true_iupac))
    except Exception:
        return False

---

Baseline benchmark results

All models are trained from scratch across 3 seeds (0, 1, 2) on a single 141 GB NVIDIA H200 GPU in BF16 and evaluated per spectrum on all 76,774 test rows:

Model	Params	Checkpoint	Inference setting	Top-1 structural accuracy	Mean NFE	Batched latency
GlycoGen-DFM + Crystallizing Flows	50M	rlacombe/glycogen-dfm-50m	K = 1	90.19% ± 0.04%	1.0	0.046 ms/spec (61×)
GlycoGen-DFM + Crystallizing Flows	50M	rlacombe/glycogen-dfm-50m	K = 8	92.51% ± 0.06%	5.0	0.201 ms/spec (13.9×)
GlycoGen-DFM + Crystallizing Flows	50M	rlacombe/glycogen-dfm-50m	K = 32	92.67% ± 0.05%	17.4	0.675 ms/spec (4.1×)
GlycoGen-DFM + Crystallizing Flows	50M	rlacombe/glycogen-dfm-50m	K = 64	92.69% ± 0.06%	33.9	1.306 ms/spec (2.1×)
GlycoGen-MDLM + Crystallizing Flows	50M	rlacombe/glycogen-mdlm-50m	K = 1	89.73% ± 0.17%	1.0	0.046 ms/spec (61×)
GlycoGen-MDLM + Crystallizing Flows	50M	rlacombe/glycogen-mdlm-50m	K = 8	92.30% ± 0.12%	5.0	0.199 ms/spec (14.1×)
GlycoGen-MDLM + Crystallizing Flows	50M	rlacombe/glycogen-mdlm-50m	K = 32	92.47% ± 0.11%	17.4	0.676 ms/spec (4.1×)
GlycoGen-AR	46.4M	rlacombe/glycogen-ar-50m	beam = 1	90.38% ± 0.05%	33.0	2.21 ms/spec (1.3×)
GlycoGen-AR	46.4M	rlacombe/glycogen-ar-50m	beam = 32 (seed 0)	90.39%	1,045	72.66 ms/spec
GlycoBART reproduction	207M	rlacombe/glycobart-207m-repro	beam = 1	80.92% ± 0.31%	23.0	2.80 ms/spec (1.0×)
GlycoBART reproduction	207M	rlacombe/glycobart-207m-repro	beam = 32 (seed 0)	78.57% (86.35% top-5)	720	91.21 ms/spec

---

Deterministic rebuild from raw Zenodo archive

You can regenerate all split files byte-for-byte on CPU (~30 minutes on an Apple M3 laptop) from the upstream Zenodo record 10.5281/zenodo.15741423 using the rlacombe/GlycoGen repository:

git clone https://github.com/rlacombe/GlycoGen.git
cd GlycoGen
pip install -r requirements.txt

# 1. Place dataset.xlsx (or converted dataset.parquet) and glycans_antennae.csv
#    from https://zenodo.org/records/15741423 into ./data/raw/

# 2. Generate row_manifest.csv, sentences_benchmark.pkl, and vocab_glycobart.txt
python data_utils/make_benchmark_split.py \
    --dataset  ./data/raw/dataset.parquet \
    --antennae ./data/raw/glycans_antennae.csv \
    --out_dir  ./data

# 3. Generate aligned X_{train,val,test}.pkl and y_{train,val,test}.pkl
python data_utils/make_aligned_numerical_pkls.py

---

License and attribution

This benchmark dataset is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license, matching the license of the upstream CandyCrunch v2 archive (Zenodo 10.5281/zenodo.15741423).

• Upstream experimental MS/MS curation and run-grouped quality filtering: James Urban, Chunsheng Jin, Kristina A. Thomsson, Niclas G. Karlsson, Callum M. Ives, Elisa Fadda, and Daniel Bojar (Nature Methods, 2024).
• Raw table archive and antenna decomposition table (glycans_antennae.csv): Ejas Althaf Abtheen, Sriram Neelamegham, and Rudiyanto Gunawan (Zenodo 10.5281/zenodo.15741423, 2025).
• Synchronized LC-MS file-grouped benchmark split and dual-channel numerical packaging: Romain Lacombe (NeurIPS 2026).

---

Citation

If you use this benchmark, please cite the original CandyCrunch v2 dataset paper, the upstream Zenodo archive, and the GlycoGen paper:

@article{candycrunch2024,
  title   = {Predicting glycan structure from tandem mass spectrometry via deep learning},
  author  = {Urban, James and Jin, Chunsheng and Thomsson, Kristina A. and
             Karlsson, Niclas G. and Ives, Callum M. and Fadda, Elisa and
             Bojar, Daniel},
  journal = {Nature Methods},
  volume  = {21},
  number  = {7},
  pages   = {1206--1215},
  year    = {2024},
  doi     = {10.1038/s41592-024-02314-6}
}

@misc{zenodo2024cc2,
  title        = {Transformer-based Deep Learning for Glycan Structure Inference from Tandem Mass Spectrometry {[Dataset]}},
  author       = {Abtheen, Ejas Althaf and Neelamegham, Sriram and Gunawan, Rudiyanto},
  howpublished = {Zenodo, \url{https://zenodo.org/records/15741423}},
  year         = {2025},
  doi          = {10.5281/zenodo.15741423}
}

@inproceedings{lacombe2026glycogen,
  title     = {GlycoGen: Glycan Structure Prediction with Crystallizing Flows},
  author    = {Lacombe, Romain},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}