Prot2Text V2 11B Instruct Hf
Pro2Text-V2: Protein Function Prediction with Multimodal Contrastive Alignment
This is the official repository for the paper "Prot2Text-V2: Protein Function Prediction with Multimodal Contrastive Alignment" by Xiao Fei, Michail Chatzianastasis, Sarah Almeida Carneiro, Hadi Abdine, Lawrence P. Petalidis, and Michalis Vazirgiannis.
We're excited to share that our paper has been accepted to NeurIPS 2025! The pretrained model weights and the dataset are now publicly available here.
š¤ Online Server š ArXiV Preprint 2505.11194 š NeurIPS 2025 Poster š» GitHub Repository š¤ Experimental Dataset
Prot2Text-V2 treats a protein sequence as if it were another language, and then translate it into English. The model takes the raw amino acid sequence as input and generates a clear, human-readable paragraph describing what the protein does.
The model is an innovative fusion of three key components:
Protein language model as sequence encoder: `facebook/esm2t363BUR50D` Modality adapter as a unique and lightweight component that bridges the gap between protein embeddings and the language model. Natural language decoder for generating articulate textual descriptions utilizing the sequence embeddings: `meta-llama/Llama-3.1-8B-Instruct`
For detailed instructions on fine-tuning the model and reproducing the experiments, please refer to our GitHub page.
If you find our research helpful, feel free to šļø cite our work or ā¤ļø like the page: