KlaBenchmark
KlaBenchmark is a benchmark release for lysine lactylation (Kla) site prediction. This Hugging Face package contains the complete benchmark TSV files under dataset/ and pretrained model checkpoints under checkpoint/.
The release is designed to be used together with the KlaBenchmark/KlaAtlas code repository, which provides the corresponding training and prediction scripts for seven Kla predictors:
ABFF-KlaAuto-KlaDeepKlaDeepKlapredHybridKlaPBertKlaPCBert-Kla
Repository Contents
1KlaBenchmark/
2├── dataset/
3│ ├── ABFF-Kla/
4│ ├── Auto-Kla/
5│ ├── DeepKla/
6│ ├── DeepKlapred/
7│ ├── HybridKla/
8│ ├── PBertKla/
9│ └── PCBert-Kla/
10└── checkpoint/
11 ├── ABFF-Kla/
12 ├── Auto-Kla/
13 ├── DeepKla/
14 ├── DeepKlapred/
15 ├── HybridKla/
16 ├── PBertKla/
17 └── PCBert-Kla/
Each method-specific folder keeps the file layout expected by the corresponding model implementation.
Dataset Structure
Each method folder in dataset/ contains benchmark partitions for different evaluation settings:
| Directory | Description |
|---|
protein_split_dataset/ | Protein-level train/validation/test split. |
sequence_split_dataset/ | Sequence-similarity-controlled split. |
15aa_window_protein_split_dataset/ | Protein-level split using 15-aa sequence windows. |
31aa_window_protein_split_dataset/ | Protein-level split using 31-aa sequence windows. |
51aa_window_protein_split_dataset/ | Protein-level split using 51-aa sequence windows. |
species_dataset/<Species>/ | Species-specific train/validation/test split. |
original_model_dataset/ | Data formatted for reproducing original model settings. |
Each partition normally contains:
1train.tsv
2val.tsv
3test.tsv
TSV Columns
Most TSV files contain the following columns:
| Column | Description |
|---|
Site | Site identifier, usually <ACC_ID>_<position>. |
ACC_ID | Protein accession ID. |
Mod_positions | Candidate lysine position in the protein. |
Species | Source organism. |
UniProtTaxID | NCBI/UniProt taxonomy identifier. |
Sequence | Local protein sequence window centered on the candidate lysine. Padding is represented by _. |
Label | 1 for experimentally identified Kla sites and 0 for background lysine sites. |
ABFF-Kla files also include method-specific structure/context fields:
| Column | Description |
|---|
Contact | Structure-derived contact window used by ABFF-Kla. |
Context | Sequence context used by ABFF-Kla. |
Dataset Sizes
The main benchmark partitions have the same size across methods:
| Partition | Train | Validation | Test |
|---|
protein_split_dataset | 380,719 | 47,604 | 47,573 |
15aa_window_protein_split_dataset | 380,719 | 47,604 | 47,573 |
31aa_window_protein_split_dataset | 380,719 | 47,604 | 47,573 |
51aa_window_protein_split_dataset | 380,719 | 47,604 | 47,573 |
sequence_split_dataset | 380,717 | 47,591 | 47,588 |
Species-specific partitions are also shared across methods:
| Species | Train | Validation | Test |
|---|
Homo_sapiens | 215,258 | 26,897 | 233,741 |
Mus_musculus | 86,174 | 10,535 | 10,706 |
Rattus_norvegicus | 36,440 | 4,525 | 4,516 |
Caenorhabditis_elegans | 18,317 | 2,291 | 2,251 |
Escherichia_coli | 13,356 | 1,691 | 1,648 |
Streptococcus_mutans | 5,160 | 656 | 624 |
The original_model_dataset/ partition is method-specific:
| Method | Train | Validation | Test |
|---|
ABFF-Kla | 12,792 | 1,422 | 3,554 |
Auto-Kla | 27,854 | 3,095 | 4,295 |
DeepKla | 3,138 | 349 | 354 |
DeepKlapred | 3,138 | 349 | 354 |
HybridKla | 33,842 | 3,761 | 9,401 |
PBertKla | 6,940 | 772 | 1,928 |
PCBert-Kla | 3,138 | 349 | 354 |
Checkpoints
The checkpoint/ directory provides pretrained artifacts for the seven supported predictors. Keep the downloaded directory structure unchanged when using the files with the companion code.
1checkpoint/<method>/
2├── preprocess.json
3└── model/
4 ├── best_model.<ext>
5 └── extra/ # Present only for methods that require side artifacts
Included checkpoint files:
| Method | Main checkpoint | Additional artifacts |
|---|
ABFF-Kla | model/best_model.keras | preprocess.json |
Auto-Kla | model/best_model.pt | preprocess.json |
DeepKla | model/best_model.keras | preprocess.json |
DeepKlapred | model/best_model.pt | preprocess.json |
HybridKla | model/best_model.pth | ESM2 state dict, feature DNNs, GPS encoder, LSTM weights, vocabulary, feature order, preprocess.json |
PBertKla | model/best_model.weights.h5 | ProteinBERT input/output encoders, preprocess.json |
PCBert-Kla | model/best_model.pt | Scaler model, preprocess.json |
Usage
Download or clone this Hugging Face repository, then point the companion code to the local dataset/ and checkpoint/ paths.
Example training command:
1python methods/<method>/train.py \
2 --train_tsv /path/to/KlaBenchmark/dataset/<method>/<partition>/train.tsv \
3 --val_tsv /path/to/KlaBenchmark/dataset/<method>/<partition>/val.tsv \
4 --test_tsv /path/to/KlaBenchmark/dataset/<method>/<partition>/test.tsv \
5 --run_name <run_name>
Example prediction command:
1python methods/<method>/predict.py \
2 --checkpoint_path /path/to/KlaBenchmark/checkpoint/<method>/model/<checkpoint_file> \
3 --input_tsv /path/to/input.tsv \
4 --output_dir /path/to/output
Some methods require additional options or external resources, such as ProteinBERT, ProtBert, ESM2, or AlphaFold structure files. See the README for the corresponding method in the companion code repository.
Label Note
Label = 0 denotes background lysine sites, not experimentally verified non-Kla sites. These residues were not identified as Kla sites in the source data and may include condition-specific or currently unobserved Kla events. Treat the benchmark as positive-versus-background classification.
Recommended Reporting
When reporting results with this release, include:
- The method name and checkpoint used.
- The exact dataset partition, for example
protein_split_dataset or species_dataset/Homo_sapiens.
- The sequence window setting when using fixed-window partitions.
- Any extra filtering, resampling, or feature preprocessing.
- Standard metrics such as AUROC, AUPRC, accuracy, F1, MCC, sensitivity, and specificity.
License and Redistribution
This release is provided for research use. Dataset redistribution and derivative use should comply with the terms of the original proteomics resources, UniProt records, pretrained backbone models, and source publications. Model weights may also inherit restrictions from their upstream architectures or training data.
Citation
If you use this benchmark, please cite the KlaBenchmark/KlaAtlas paper and the original publications for the individual predictors or pretrained protein language models used in your experiments.
