OPI-Struc

OPI-Struc: Open Protein Instructions for Structures Links **GitHub:** https://github.com/ocx-lab/STELLA **Paper:** STELLA: A Multimodal LLM for Protein Functional Annotation via Unified Sequence-Structure Encoding — accepted as a Findings Paper at ACL 2026. **Related dataset:** OPI (Open Protein Instructions) Dataset Overview OPI-Struc (Open Protein Instructions for Structures)** is a multimodal instruction-tuning dataset specifically designed for the STELLA project. It extends the OPI paradigm by integrating **protein 3D structure information** with textual instructions, enabling LLMs to perform protein-related tasks grounded in both sequence and structural modalities. OPI-Struc focuses on two critical protein-related tasks: **Functional Description Prediction (FP)**: Predicting the biological function of a protein from its 3D structure, in both free-text QA (FTQA) and multiple-choice QA (MCQA) formats. **Enzyme-catalyzed Reaction Prediction (EP)**: Predicting the enzyme name or EC number from a protein's 3D structure. Total dataset size: 351,183 training samples and 40,993 testing samples.** **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 below. Dataset Construction The OPI-Struc dataset is curated from two primary sources: **Function (FP) tasks**: Protein entries are sourced from UniProtKB/Swiss-Prot (release 2022_04), following the Prot2Text data split. 3D structures are obtained from the AlphaFold Protein Structure Database (AFDB). **Enzyme (EP) tasks**: Protein entries are sourced from the Enzyme Commission dataset, with 3D structures from RCSB PDB. Each sample is formatted as a multi-turn conversation (instruction tuning format) with a ` ` token indicating where the protein structure e