ESM3-large Protein Representations for protein-protein interaction prediction

This data repository contains the protein embeddings from the ESM3-large (98B) protein language model trained by [1] on the following datasets used for Protein-Protein Interaction site prediction tasks, in Pytorch format. The original datasets are also included in csv format, with protein ID extracted either from the Uniprot or PDB database, protein sequence and labels (0 or 1 corresponding to non-interacting or interacting residues). The following datasets are included: Biolip (training and validation partition), curated from [2]. PDBbind (training and validation partition), described in more detail in [...]. ZK448 (used for testing), curated from [3]. The original datasets taken from [2,3] were filtered to avoid similar sequences in training and testing sets. These were used to train the protein-protein interaction site prediction models described in [...]. The code for this project can be found on Github. The paper is available on .... Trained models are available from huggingface. Version changes v2: Labels in the pdbBind datasets use the van der Waals radius for the cut-off. Terms of Use Embeddings generated using the ESM3 98B model via EvolutionaryScale Forge API. Attribution: EvolutionaryScale, PBC. Non-commercial use only. Must not be used to train or improve competing protein language models. Any redistribution must maintain this attribution. References Hayes, T., Rao, R., Akin, H., Sofroniew, N. J., Oktay, D., Lin, Z., ... & Rives, A. (2025). Simulating 500 million years of evolution with a language model. Science, 387(6736), 850-858. Stringer, B., de Ferrante, H., Abeln, S., Heringa, J., Feenstra, K. A., & Haydarlou, R. (2022). PIPENN: protein interface prediction from sequence with an ensemble of neural nets. Bioinformatics, 38(8), 2111-2118. Zhang, J., & Kurgan, L. (2018). Review and comparative assessment of sequence-based predictors of protein-binding residues. Briefings in bioinformatics, 19(5), 821-837.