init

README.md

@@ -1,3 +1,70 @@
----
-license: mit
----
+---
+license: mit
+---
+
+# Dataset Card for NovoBench
+Datasets used for the baseline comparison of deep learning-based de novo peptide sequencing method
+
+## Dataset Description
+
+- **Repository:** [NovoBench](https://github.com/jingbo02/NovoBench)
+- **Paper:** [NovoBench: Benchmarking Deep Learning-based De Novo Peptide Sequencing Methods in Proteomics](https://arxiv.org/pdf/2406.11906) 
+
+
+### Dataset Summary
+
+- **Seven-species Dataset.** The Seven-species dataset contains low-resolution mass spectrum and
+ their peptide labels from 7 different species. The previous work DeepNovo has evaluated its
+ performance on these datasets with the leave-one-out method, i.e., training the model on 6 species and
+ testing on the left one species, to mimic the real-world challenging cases where we have to identify
+ the never-before-seen peptide sequences for the observed mass spectrum. In this paper, we conducted
+ testing on the yeast species and training on the remaining 6 species.
+
+- **Nine-species Dataset.** The Nine-species dataset is the most widely-used dataset by previous works
+ such as DeepNovo, PointNovo, and Casanovo, which contains high-resolution mass
+ spectrum and their peptide labels from 9 different species. We adopt the Nine-species dataset
+ used by the original publication of DeepNovo (MassIVE dataset identifier: MSV000081382) for
+ benchmarking. Similar to Seven-species dataset, we train models on 8 species and evaluate the left
+ yeast species. Additionally, these datasets contain 3 PTMs (oxidation of methionine, deamidation of
+ asparagine or glutamine), enabling the fair evaluation of various models’ performance in terms of
+ identifying PTMs.
+ - **HC-PT Dataset.** The HC-PT dataset, as detailed in the InstaNovo paper, includes synthetic tryptic
+ peptides that span all canonical human proteins and isoforms. It also encompasses peptides generated
+ by alternative proteases and HLA peptides. The key feature of the HC-PT dataset is its high-resolution
+ spectrum for human-origin peptides and the peptide labels are derived from the high-confidence
+ search results of MaxQuant.
+
+
+## Dataset Structure
+
+The dataset is tabular, where each row corresponds to a labelled MS2 spectra.
+- `sequence (string)` \
+  The target peptide sequence excluding post-translational modifications
+- `modified_sequence (string)` \
+  The target peptide sequence including post-translational modifications
+- `precursor_mz (float64)` \
+  The mass-to-charge of the precursor (from MS1)
+- `charge (int64)` \
+  The charge of the precursor (from MS1)
+- `mz_array (list[float64])` \
+  The mass-to-charge values of the MS2 spectrum
+- `mz_array (list[float32])` \
+  The intensity values of the MS2 spectrum
+
+
+
+## Citation Information
+
+If you use this dataset, please cite the NovoBench:
+
+```bibtex
+@misc{zhou2024novobenchbenchmarkingdeeplearningbased,
+      title={NovoBench: Benchmarking Deep Learning-based De Novo Peptide Sequencing Methods in Proteomics}, 
+      author={Jingbo Zhou and Shaorong Chen and Jun Xia and Sizhe Liu and Tianze Ling and Wenjie Du and Yue Liu and Jianwei Yin and Stan Z. Li},
+      year={2024},
+      eprint={2406.11906},
+      archivePrefix={arXiv},
+      primaryClass={q-bio.QM},
+      url={https://arxiv.org/abs/2406.11906}, 
+}
+```