--- extra_gated_heading: Acknowledge license to accept the repository extra_gated_prompt: > The Beijing Academy of Artificial Intelligence (hereinafter referred to as "we" or "BAAI") provides you with an open-source dataset (hereinafter referred to as "dataset") through the OPI-Struc HuggingFace repository (https://huggingface.co/datasets/BAAI/OPI-Struc). You can download the dataset you need and use it for purposes such as learning and research while abiding by the usage rules of each original dataset. Before you acquire the open-source dataset (including but not limited to accessing, downloading, copying, distributing, using, or any other handling of the dataset), you should read and understand this "OPI-Struc Open-Source Dataset Usage Notice and Disclaimer" (hereinafter referred to as "this statement"). Once you acquire the open-source dataset, regardless of your method of acquisition, your actions will be regarded as acknowledgment of the full content of this statement. 1. Ownership and Operation Rights You should fully understand that the ownership and operation rights of the OPI-Struc HuggingFace repository (including the current and all previous versions) belong to BAAI. BAAI has the final interpretation and decision rights over this platform/tool and the open-source dataset plan. You acknowledge and understand that due to updates and improvements in relevant laws and regulations and the need to fulfill our legal compliance obligations, we reserve the right to update, maintain, or even suspend or permanently terminate the services of this platform/tool from time to time. We will notify you of possible situations mentioned above reasonably such as through an announcement or email within a reasonable time. You should make corresponding adjustments and arrangements in a timely manner. However, we do not bear any responsibility for any losses caused to you by any of the aforementioned situations. 2. Claim of Rights to Open-Source Datasets For the purpose of facilitating your dataset acquisition and use for learning, and research, we have performed necessary steps such as format integration, data cleaning, labeling, categorizing, annotating, and other related processing on the third-party original datasets to form the open-source datasets for this platform/tool's users. You understand and acknowledge that we do not claim the proprietary rights of intellectual property to the open-source datasets. Therefore, we have no obligation to actively recognize and protect the potential intellectual property of the open-source datasets. However, this does not mean that we renounce the personal rights to claim credit, publication, modification, and protection of the integrity of the work (if any) of the open-source datasets. The potential intellectual property and corresponding legal rights of the original datasets belong to the original rights holders. In addition, providing you with open-source datasets that have been reasonably arranged, processed, and handled does not mean that we acknowledge the authenticity, accuracy, or indisputability of the intellectual property and information content of the original datasets. You should filter and carefully discern the open-source datasets you choose to use. You understand and agree that BAAI does not undertake any obligation or warranty responsibility for any defects or flaws in the original datasets you choose to use. 3. Usage Restrictions for Open-Source Datasets Your use of the dataset must not infringe on our or any third party's legal rights and interests (including but not limited to copyrights, patent rights, trademark rights, and other intellectual property and other rights). After obtaining the open-source dataset, you should ensure that your use of the open-source dataset does not exceed the usage rules explicitly stipulated by the rights holders of the original dataset in the form of a public notice or agreement, including the range, purpose, and lawful purposes of the use of the original data. We kindly remind you here that if your use of the open-source dataset exceeds the predetermined range and purpose of the original dataset, you may face the risk of infringing on the legal rights and interests of the rights holders of the original dataset, such as intellectual property, and may bear corresponding legal responsibilities. 4. Personal Information Protection Due to technical limitations and the public welfare nature of the open-source datasets, we cannot guarantee that the open-source datasets do not contain any personal information, and we do not bear any legal responsibility for any personal information that may be involved in the open-source datasets. If the open-source dataset involves personal information, we do not bear any legal responsibility for any personal information processing activities you may involve when using the open-source dataset. We kindly remind you here that you should handle personal information in accordance with the provisions of the "Personal Information Protection Law" and other relevant laws and regulations. To protect the legal rights and interests of the information subject and to fulfill possible applicable laws and administrative regulations, if you find content that involves or may involve personal information during the use of the open-source dataset, you should immediately stop using the part of the dataset that involves personal information and contact us as indicated in "6. Complaints and Notices." 5. Information Content Management We do not bear any legal responsibility for any illegal and bad information that may be involved in the open-source dataset. If you find that the open-source dataset involves or may involve any illegal and bad information during your use, you should immediately stop using the part of the dataset that involves illegal and bad information and contact us in a timely manner as indicated in "6. Complaints and Notices." 6. Complaints and Notices If you believe that the open-source dataset has infringed on your legal rights and interests, you can contact us at 010-50955974, and we will handle your claims and complaints in accordance with the law in a timely manner. To handle your claims and complaints, we may need you to provide contact information, infringement proof materials, and identity proof materials. Please note that if you maliciously complain or make false statements, you will bear all legal responsibilities caused thereby (including but not limited to reasonable compensation costs). 7. Disclaimer You understand and agree that due to the nature of the open-source dataset, the dataset may contain data from different sources and contributors, and the authenticity, accuracy, and objectivity of the data may vary, and we cannot make any promises about the availability and reliability of any dataset. In any case, we do not bear any legal responsibility for any risks such as personal information infringement, illegal and bad information dissemination, and intellectual property infringement that may exist in the open-source dataset. In any case, we do not bear any legal responsibility for any loss (including but not limited to direct loss, indirect loss, and loss of potential benefits) you suffer or is related to the open-source dataset. 8. Others The open-source dataset is in a constant state of development and change. We may update, adjust the range of the open-source dataset we provide, or suspend, pause, or terminate the open-source dataset service due to business development, third-party cooperation, changes in laws and regulations, and other reasons. extra_gated_fields: Name: text Affiliation: text Country: text I agree to accept the license: checkbox extra_gated_button_content: Acknowledge license license: cc-by-nc-4.0 task_categories: - text-generation - question-answering - multiple-choice language: - en tags: - Biology - Protein - Multimodal LLM - AI for Life Science pretty_name: Open Protein Instructions for Structures (OPI-Struc) size_categories: - 100K **Note:** This repository provides **annotation JSON files only**. The protein structure embedding files (`embs_pt/`) are **not** included due to their large size. Each `embs_pt/` directory contains a placeholder. Users should generate embeddings locally using the provided script — see [Generating Protein Structure Embeddings](#generating-protein-structure-embeddings) below. ## Dataset Construction The OPI-Struc dataset is curated from two primary sources: - **Function (FP) tasks**: Protein entries are sourced from [UniProtKB/Swiss-Prot](https://www.uniprot.org/uniprotkb?facets=reviewed%3Atrue&query=%2A) (release 2022_04), following the [Prot2Text](https://ojs.aaai.org/index.php/AAAI/article/view/28948) data split. 3D structures are obtained from the [AlphaFold Protein Structure Database (AFDB)](https://alphafold.ebi.ac.uk/). - **Enzyme (EP) tasks**: Protein entries are sourced from the [Enzyme Commission dataset](https://www.enzyme-database.org/), with 3D structures from [RCSB PDB](https://www.rcsb.org/). Each sample is formatted as a multi-turn conversation (instruction tuning format) with a `` token indicating where the protein structure embedding should be inserted. ## Dataset Statistics | Task | Training Set | Training Size | Testing Set | Testing Size | Metrics | Protein Source | |------|-------------|---------------|-------------|-------------|---------|----------------| | FPFTQA | Functiontrain\_FTQA (+aug) | 248,315 (+49,663) | Functiontest\_FTQA
Functiontest\_FTQA\_v2401
Functiontest\_FTQA\_trunc90 | 4,203
270
4,203 | BLEU-4, BERT-score, ROUGE | AFDB | | FPMCQA | Functiontrain\_MCQA | 24,000 | Functiontest\_MCQA\_1X
Functiontest\_MCQA\_4X | 4,203
16,812 | Accuracy | AFDB | | EP | Enzymetrain | 29,205 | Enzymetest
Enzymetest\_EC\_number | 5,651
5,651 | Accuracy | PDB | - **Functiontest\_FTQA\_v2401**: A temporal out-of-distribution test set constructed from Swiss-Prot release 2024_01, used to evaluate zero-shot generalization on unseen proteins. - **Functiontest\_FTQA\_trunc90**: A structural degradation test set where protein structures are truncated to 90% of their original residues, used to evaluate robustness to incomplete structures. - **Functiontest\_MCQA\_1X** vs. **Functiontest\_MCQA\_4X**: The 1X version has options without permutation; the 4X version has options with permutation (4 orderings per question). Note: FTQA - Free-Text Question Answering, MCQA - Multi-Choice Question Answering ## Dataset Folder Structure This repository provides annotation JSON files and `embs_pt/` placeholder directories. The folder structure is organized by **protein encoder** (ESM3, Prot2Text, SaProt) since each encoder produces different embeddings. ``` OPI-Struc/ ├── Function/ │ ├── esm3/ │ │ ├── train/ │ │ │ ├── ann.json # FP_FTQA training (248,315 samples) │ │ │ ├── function_aug_49663.json # FP_FTQA augmented training (49,663 samples) │ │ │ ├── ann_multichoice_24k.json # FP_MCQA training (24,000 samples) │ │ │ └── embs_pt/ # [placeholder] ESM3 embeddings │ │ ├── test/ │ │ │ ├── ann.json # FP_FTQA test (4,203 samples) │ │ │ ├── ann_multichoice_1x.json # FP_MCQA_1X test (4,203 samples) │ │ │ ├── ann_multichoice_4x.json # FP_MCQA_4X test (16,812 samples) │ │ │ └── embs_pt/ # [placeholder] ESM3 embeddings │ │ ├── test_SwissProt_v2401/ │ │ │ ├── ann.json # FP_FTQA temporal OOD test (270 samples) │ │ │ └── embs_pt/ # [placeholder] ESM3 embeddings │ │ └── test_struct_trunc90/ │ │ ├── ann.json # FP_FTQA structural degradation test (4,203 samples) │ │ └── embs_pt/ # [placeholder] ESM3 embeddings │ ├── Prot2Text/ │ │ ├── train/ │ │ │ ├── ann.json # Same annotations as esm3/train/ann.json │ │ │ └── embs_pt/ # [placeholder] Prot2Text embeddings │ │ └── test/ │ │ └── embs_pt/ # [placeholder] Prot2Text embeddings │ └── SaProt_repr/ │ ├── train/ │ │ ├── ann.json # Same annotations as esm3/train/ann.json │ │ └── embs_pt/ # [placeholder] SaProt embeddings │ └── test/ │ └── embs_pt/ # [placeholder] SaProt embeddings │ ├── Enzyme/ │ ├── esm3/ │ │ ├── train/ │ │ │ ├── ann.json # EP training (29,205 samples) │ │ │ └── embs_pt/ # [placeholder] ESM3 embeddings │ │ └── test/ │ │ ├── ann.json # EP test (5,651 samples) │ │ ├── ann_ec_number.json # EP EC number test (5,651 samples) │ │ └── embs_pt/ # [placeholder] ESM3 embeddings │ ├── Prot2Text/ │ │ ├── train/ │ │ │ └── embs_pt/ # [placeholder] Prot2Text embeddings │ │ └── test/ │ │ └── embs_pt/ # [placeholder] Prot2Text embeddings │ └── ... │ └── README.md # This file ``` > **Note:** The annotation JSON files are **identical** across encoder directories for the same task and split. They are duplicated in each encoder's directory to match the expected layout of STELLA's dataloader, which looks for `ann.json` and `embs_pt/` in the same parent directory. ## Data Format ### FP (Functional Description Prediction) — FTQA Format Each entry follows the conversation format used by [LLaVA](https://github.com/haotian-liu/LLaVA): ```json { "swissprot_id": "P07412", "sequence": "GFLTAEEKGLVNGLWGKVNVDEVGGEALGRLLVVYPWTQRFFESFGDLSS...", "AFDB_id": "AF-P07412-F1-model_v4", "conversations": [ { "from": "human", "value": "\nCan you furnish a comprehensive description outlining the function associated with the protein?" }, { "from": "gpt", "value": "Involved in oxygen transport from the lung to the various peripheral tissues." } ] } ``` ### FP — MCQA Format ```json { "swissprot_id": "P62877", "sequence": "MAAAMDVDTPSGTNSGAGKKRFEVKKWNAVALWAWDIVVDNCAICRNHIM...", "AFDB_id": "AF-P62877-F1-model_v4", "conversations": [ { "from": "human", "value": "\nWhat are the main functions of this protein?\nA. [option A text]\nB. [option B text]\nC. [option C text]\nD. [option D text]" }, { "from": "gpt", "value": "A" } ] } ``` ### EP (Enzyme-catalyzed Reaction Prediction) Format ```json { "PDB_id": "2pmo.X", "conversations": [ { "from": "human", "value": "\nPlease share the enzyme terminology for this protein." }, { "from": "gpt", "value": "non-specific serine/threonine protein kinase" } ] } ``` ### EP — EC Number Format (test only) ```json { "PDB_id": "4xi6.A", "EC_number": "2.3.2.27" } ``` ### Key Fields | Field | Description | |-------|-------------| | `swissprot_id` | UniProtKB/Swiss-Prot accession ID | | `AFDB_id` | AlphaFold DB structure identifier (e.g., `AF-P07412-F1-model_v4`) | | `PDB_id` | PDB structure identifier with chain (e.g., `2pmo.X`) | | `sequence` | Amino acid sequence of the protein | | `conversations` | Multi-turn conversation in LLaVA format; `` marks where the structure embedding is injected | | `EC_number` | Enzyme Commission classification number | ## Generating Protein Structure Embeddings Since the pre-computed embedding files (`embs_pt/`) are too large to host on Hugging Face, you need to generate them locally before training or evaluating STELLA. ### Prerequisites 1. **Download raw protein structure files**: - **For FP tasks** (SwissProt/AFDB data): Download AlphaFold DB structures (`.pdb` files). Each protein's structure ID is in the `AFDB_id` field (e.g., `AF-P07412-F1-model_v4`). Download from the [AlphaFold Protein Structure Database](https://alphafold.ebi.ac.uk/). - **For EP tasks** (Enzyme/PDB data): Download PDB structures. Each protein's structure ID is in the `PDB_id` field (e.g., `2pmo.X`). Download from [RCSB PDB](https://www.rcsb.org/). 2. **Set up the STELLA environment** following the [STELLA installation guide](https://github.com/ocx-lab/STELLA). 3. **Download protein encoder checkpoints**: - [ESM3 (esm3-sm-open-v1)](https://huggingface.co/EvolutionaryScale/esm3-sm-open-v1) - [Prot2Text (prot2text_large)](https://huggingface.co/hadi-abdine/Prot2Text) - [SaProt (SaProt_650M_AF2)](https://huggingface.co/westlake-repl/SaProt_650M_AF2) ### Running the Embedding Generation Script Use [scripts/precompute_embeddings.py](https://github.com/ocx-lab/STELLA/blob/main/scripts/precompute_embeddings.py) from the STELLA repository: ```bash # ESM3 embeddings for Function training data python scripts/precompute_embeddings.py \ --encoder esm3 \ --ann_json /path/to/OPI-Struc/Function/esm3/train/ann.json \ --structure_dir /path/to/alphafold_pdb_files/ \ --output_dir /path/to/OPI-Struc/Function/esm3/train/embs_pt/ \ --encoder_path /path/to/esm3-sm-open-v1 # Prot2Text embeddings for Function training data python scripts/precompute_embeddings.py \ --encoder prot2text \ --ann_json /path/to/OPI-Struc/Function/Prot2Text/train/ann.json \ --structure_dir /path/to/alphafold_pdb_files/ \ --output_dir /path/to/OPI-Struc/Function/Prot2Text/train/embs_pt/ \ --encoder_path /path/to/prot2text_large # SaProt embeddings for Function training data python scripts/precompute_embeddings.py \ --encoder saprot \ --ann_json /path/to/OPI-Struc/Function/SaProt_repr/train/ann.json \ --structure_dir /path/to/alphafold_pdb_files/ \ --output_dir /path/to/OPI-Struc/Function/SaProt_repr/train/embs_pt/ \ --encoder_path /path/to/SaProt_650M_AF2/SaProt_650M_AF2.pt \ --foldseek_path stella/model/multimodal_encoder/bin/foldseek # ESM3 embeddings for Enzyme training data python scripts/precompute_embeddings.py \ --encoder esm3 \ --ann_json /path/to/OPI-Struc/Enzyme/esm3/train/ann.json \ --structure_dir /path/to/pdb_files/ \ --output_dir /path/to/OPI-Struc/Enzyme/esm3/train/embs_pt/ \ --encoder_path /path/to/esm3-sm-open-v1 ``` > **Notes:** > - The script supports **resume**: it automatically skips proteins whose `.pt` files already exist in the output directory. > - Output `embs_pt/` directories must be placed alongside the corresponding `ann.json` file, as STELLA's dataloader expects this layout. > - For **SaProt**, the `--foldseek_path` argument is required. > - For **Prot2Text**, [graphein](https://github.com/a-r-j/graphein) and DSSP 3.0 must be installed. ### Embedding Output Format Each protein produces a single `.pt` file named `{structure_id}.pt`: - **ESM3**: tensor of shape `[1, L, 1536]` (per-residue embeddings, L = sequence length) - **Prot2Text**: dict with `hidden_states` (tensor `[1, 1021, 768]`) and `attentions` (tensor `[1, 1021]`) - **SaProt**: tensor of shape `[1, L, 1280]` (per-residue representations) ## Data Sources | Source | URL | Usage | |--------|-----|-------| | UniProtKB/Swiss-Prot (release 2022_04) | [https://www.uniprot.org/](https://www.uniprot.org/) | FP task protein annotations | | UniProtKB/Swiss-Prot (release 2024_01) | [https://www.uniprot.org/](https://www.uniprot.org/) | Temporal OOD test set (v2401) | | AlphaFold Protein Structure Database | [https://alphafold.ebi.ac.uk/](https://alphafold.ebi.ac.uk/) | FP task protein 3D structures | | RCSB Protein Data Bank | [https://www.rcsb.org/](https://www.rcsb.org/) | EP task protein 3D structures | | Enzyme Commission Database | [https://www.enzyme-database.org/](https://www.enzyme-database.org/) | EP task enzyme annotations | ## Related Datasets - [OPI (Open Protein Instructions)](https://huggingface.co/datasets/BAAI/OPI) — The sequence-only instruction dataset for adapting LLMs to protein tasks (NeurIPS 2024 Workshop). ## Citation If you use OPI-Struc in your research, please cite: ```bibtex @inproceedings{stella2026, title={STELLA: A Multimodal LLM for Protein Functional Annotation via Unified Sequence-Structure Encoding}, author={Hongwang Xiao, Wenjun Lin, Xi Chen, Hui Wang, Kai Chen, Jiashan Li, Yuancheng Sun, Sicheng Dai, Boya Wu, Qiwei Ye}, booktitle={Findings of the Association for Computational Linguistics: ACL 2026}, year={2026} } ``` ## License This dataset is licensed under [Creative Commons Attribution Non Commercial 4.0 (CC BY-NC 4.0)](https://creativecommons.org/licenses/by-nc/4.0/). The use of this dataset must also comply with the original licenses and terms of the upstream data sources: - [UniProt License & Disclaimer](https://www.uniprot.org/help/license) - [AlphaFold DB Terms of Use](https://www.ebi.ac.uk/about/terms-of-use) - [RCSB PDB Usage Policy](https://www.rcsb.org/pages/usage-policy) - [ESM3 Community License](https://github.com/evolutionaryscale/esm/blob/main/LICENSE.md) (for ESM3 embeddings) - [Prot2Text License](https://github.com/hadi-abdine/Prot2Text/blob/master/LICENSE) (for Prot2Text embeddings) - [SaProt License](https://github.com/westlake-repl/SaProt/blob/main/LICENSE) (for SaProt embeddings)