Protap

• Protap
  • Overview
  • Installation
  • Dataset Description
  • Citation
  • Contact

Overview

Protap is a benchmark dataset for evaluating protein modeling algorithms in five biologically realistic downstream applications. It enables comparative evaluation of both pre-trained models and domain-specific architectures. Protap includes both sequence and structural data, with tasks ranging from protein function prediction to targeted protein degradation.

(I) Masked Language Modeling(MLM) is a self-supervised objective designed to recover masked residues in protein sequences; (II) Multi-View Contrastive Learning(MVCL) leverages protein structural information by aligning representations of biologically correlated substructures. (III) Protein Family Prediction(PFP) introduces functional and structural supervision by training models to predict family labels based on protein sequences and 3D structures.

Protap was introduced in the paper
Protap: A Benchmark for Protein Modeling on Realistic Downstream Applications
by Yan et al., arXiv 2025.

Installation

git clone https://github.com/Trust-App-AI-Lab/protap.git
cd protap

Dataset-Description

📦 Configurations

Config Name	Task Description	Modality	Split	File Types
AFP	Protein function annotation (GO term prediction)	Seq + Struct	test	.csv,json
PCSP	Cleavage site prediction (enzyme-substrate pairs)	Seq + Struct	train/test	.pkl
PLI_DAVIS	Protein–ligand binding affinity regression	Seq + Struct	test	.txt, .json, .sdf
PROTACs	Ternary complex prediction in targeted degradation	Seq + Struct	test	.txt, .json, .sdf

---

💡 Task Descriptions

🔹 Enzyme-Catalyzed Protein Cleavage Site Prediction (PCSP)

• Description: Predict residue-level cleavage sites under the catalysis of enzymes.
• Input: A protein substrate and an enzyme, both represented by sequences and 3D structures.
• Output: A binary vector indicating whether each residue is a cleavage site.
• Files:
  • Train: PCSP/train_C14005.pkl, PCSP/train_M10003.pkl
  • Test: PCSP/test_C14005.pkl, PCSP/test_M10003.pkl
• Format:
  • Input Format:
    Python pickle file (.pkl) with a list of samples. Each sample is a dictionary with keys like:
    • enzyme_seq: amino acid string
    • enzyme_coords: array of 3D coordinates
    • substrate_seq: amino acid string
    • substrate_coords: array of 3D coordinates
    • Structural data of proteins:

        {
        "Q5QJ38": {
          "name": "Q5QJ38",
          "seq": "MPQLLRNVLCVIETFHKYASEDSNGAT...",
          "coords": [[[-0.432, 25.507, -8.242], ...], ...],
          "cleave_site": [136]
        }
      }
    

    -Sequence and structure of substrate proteins

    {
    "P31001_MER0000622": [110, 263],
    "000232_MER0000622": [276, 334, 19],
    ...
    }
    

  • Label Format:
    cleavage_sites: list of 0/1 values (length = number of substrate residues)
• Metric: AUC, AUPR

---

🔹 Targeted Protein Degradation by Proteolysis-Targeting Chimeras (PROTACs)

• Description: Predict whether a given PROTAC, E3 ligase, and target protein form a functional ternary complex for degradation.
• Input:
  • PROTAC molecule (warhead, linker, E3 ligand), target protein, and E3 ligase
• Output: Binary label (1: degradation occurs, 0: no degradation)
• Files:
  • PROTACs/PROTAC_clean_structure_label.txt
  • PROTACs/protac_poi_e3ligase_structure.json
  • Subfolders: PROTACs/e3_ligand/, linker/, warhead/
• Format:

  • Input Format:

    • *.json: Contains structural coordinates of proteins
    • *.sdf or .mol under folders: 3D conformers of molecules (SMILES-based)
    • PROTAC_clean_structure_label.txt: sample list with molecule paths and label

          [
        "Q8IWV7": {
          "seq": "MADEEAGGTERMEISAELPQTPQRLASWWDQQVDFYTA...",
          "coord": [[21.9960, 68.3170, -49.9029], ...]
        },
        .....
      ]
    

      uniprot    e3_ligase_structure    linker_sdf      warhead_sdf     e3_ligand_sdf     label
      Q00987     Q96SW2                 linker_2.sdf    warhead_7.sdf   e3_ligand_7.sdf   1
      Q00987     Q96SW2                 linker_2.sdf    warhead_7.sdf   e3_ligand_16.sdf  1
      P10275     P40337                 linker_13.sdf   warhead_27.sdf  e3_ligand_27.sdf  0
      P10275     P40337                 linker_33.sdf   warhead_27.sdf  e3_ligand_27.sdf  0
    

  • Label Format:
    Binary label (0 or 1) in final column of *_label.txt

• Metric: Accuracy, AUC

---

🔹 Protein–Ligand Interactions (PLI_DAVIS)

• Description: Predict the binding affinity between a protein and a small molecule ligand.

• Input: Protein and ligand 3D structures

• Output: Real-valued regression target representing binding affinity

• Files:

  • PLI_DAVIS/davis_drug_pdb_data.txt
  • PLI_DAVIS/pli_structure.json
  • PLI_DAVIS/data/: contains structure files for small molecules

• Format:

  • Input Format:
    • pli_structure.json: Dictionary of protein structures with residue positions
    • davis_drug_pdb_data.txt: Tab-separated file with fields: drug, protein, y_true
    • data/: ligand structures (.sdf, .pdbqt or similar)

      [
        "4WSQ.B": {
          "uniprot_id": "Q2M2I8",
          "seq": "EVLAEGGFAIVFLCALKRMVCKREIQIMRDLS...",
          "coord": [[[6.6065, 16.2524, 52.3289], ...], ...]
        }
      ]
  

    drug    protein    Kd        y        protein_pdb
    5291    AAK1       10000.0   5.0      4WSQ.B
    5291    ABL1p      10000.0   5.0      3QRJ.B
    5291    ABL2       10.0      7.99568  2XYN.C
  

  • Label Format: Real-valued log-transformed binding affinity (e.g., pKd or −log(Kd))

• Metric: Mean Squared Error (MSE), Pearson Correlation

---

🔹 Protein Function Annotation Prediction (AFP)

• Description: Predict GO (Gene Ontology) terms for proteins.
• Input: Protein sequence and structure
• Output: Multi-label vector representing associated GO terms
• Files:
  • AFP/processed_data.json
  • AFP/data_splits.json
  • AFP/nrPDB-GO_annot.tsv
  • AFP/nrPDB-GO_test.csv
• Format:
  • Input Format: AFP/processed_data.json: json file with key value pair:
    • name: unique ID
    • seq: amino acid sequence
    • coords: protein structure information
    • molecular_function,biological_process,cellular_component: GO term labels

    - [
      {
        "name": "2P1Z-A",
        "seq": "SKKAELAELVKELAVYVDLRRATLHARASRLIGELLRELTADWDYVA...",
        "coords": [[ [6.4359, 51.3870, 15.4490], ... ], ...],
        "molecular_function": [0, 0, ..., 1, ...],
        "biological_process": [0, 0, ..., 0, ...],
        "cellular_component": [0, 0, ..., 1, ...]
      },
      ...
    ]
    

    AFP/data_splits.json: The division of the train and test datasets

  • annotation file Format:
    AFP/nrPDB-GO_annot.tsv：tsv file with multi-column contains the detailed GO term information for all proteins.

    ### PDB-chain	GO-terms (molecular_function)	GO-terms (biological_process)	GO-terms (cellular_component)
    1A6X-A	        GO:0016874	                    GO:0016053, ... , GO:0032787    GO:1902494
    

    AFP/nrPDB-GO_test.csv: csv file with multi-column contains the sequence similarity between the test proteins and the training set.

    PDB-chain,<30%,<40%,<50%,<70%,<95%
    3ONG-B   ,  0 ,  1 ,  1 , 1  , 1
    1AD3-A   ,  0 ,  0 ,  1 , 1  , 1
    1FI8-A   ,  0 ,  0 ,  0 , 1  , 1
    2FWM-X   ,  0 ,  1 ,  1 , 1  , 1
    

  • Label Format: One or more GO term IDs (e.g., GO:0007165) in a multi-hot encoded label vector, The detailed GO term label information for each protein can be found in the AFP/nrPDB-GO_annot.tsv file.
• Metric: Fmax, AUPR

Citation

If you use this dataset, please cite:

  @misc{yan2025protapbenchmarkproteinmodeling,
        title={Protap: A Benchmark for Protein Modeling on Realistic Downstream Applications}, 
        author={Shuo Yan and Yuliang Yan and Bin Ma and Chenao Li and Haochun Tang and Jiahua Lu and Minhua Lin and Yuyuan Feng and Hui Xiong and Enyan Dai},
        year={2025},
        eprint={2506.02052},
        archivePrefix={arXiv},
        primaryClass={q-bio.BM},
        url={https://arxiv.org/abs/2506.02052}, 
  }

Contact

Please submit GitHub issues or contact Shuo Yan (syan225@connect.hkust-gz.edu.cn ) for any questions related to the source code.