Data for "SeaMoon: from protein language models to continuous structural heterogeneity"

Datasets used for development of SeaMoon: https://github.com/PhyloSofS-Team/seamoon.

This upload contains the following data:

precomputed_emb.tar.gz is a compressed archive containing the precomputed data used for training and testing the models of the SeaMoon method, in Torch .pt format. The file prefixes consist of two IDs, "ID1_ID2_", identifying the DANCE [1] protein conformational collection used for its generation. "ID1" represents the first member of the collection in alphabetical order, while "ID2" is the reference conformation for the structural alignment. The "ESM_data" or "ProstT5_data" suffixes designate the type of embeddings, generated by either ESM2 [2] or ProstT5 [3].The dictionnary contains the following keys:

emb: The per-residue embedding.

data: A tuple containing "ID2" (the reference), the amino acid sequence, and the coverage of the positions in the original DANCE collection.

eigvect: The eigenvectors of the covariance matrix of the "ID1_ID2" collection, centered on reference conformaton "D2".

eigval: The associated eigenvalues.

ref: The coordinates of the C-alpha atoms of the reference conformaton "ID2".

train_list.txt, train_list_5ref.txt, val_list.txt and test_list.txt contain the identifiers of the samples used for training and evaluating the SeaMoon models. In the "5ref" setting, we used up to 5 reference conformations per collection. 

For details on SeaMoon see:

SeaMoon: Prediction of molecular motions based on language models

Valentin Lombard, Dan Timsit, Sergei Grudinin, Elodie Laine

bioRxiv 2024.09.23.614585; doi: https://doi.org/10.1101/2024.09.23.614585

For more information on data usage and generation please see https://github.com/PhyloSofS-Team/seamoon.

Abstract:

How protein move and deform determines their interactions with the environment and is thus of utmost importance for cellular functioning. Following the revolution in single protein 3D structure prediction, researchers have focused on repurposing or developing deep learning models for sampling alternative protein conformations. In this work, we explored whether continuous compact representations of protein motions could be predicted directly from protein sequences, without exploiting nor sampling protein structures. Our approach, called SeaMoon, leverages protein Language Model (pLM) embeddings as input to a lightweight (~1M trainable parameters) convolutional neural network. SeaMoon achieves a success rate of up to 40% when assessed against ~1,000 collections of experimental conformations exhibiting a wide range of motions. SeaMoon capture motions not accessible to the normal mode analysis, an unsupervised physics-based method relying solely on a protein structure's 3D geometry, and generalises to proteins that do not have any detectable sequence similarity to the training set. SeaMoon is easily retrainable with novel or updated pLMs. 

[1] Lombard, V.; Grudinin, S.; Laine, E. Explaining Conformational Diversity in Protein Families through Molecular Motions. Scientific Data 2024, 11, 752.

[2] Lin, Z.; Akin, H.; Rao, R.; Hie, B.; Zhu, Z.; Lu, W.; Smetanin, N.; Verkuil, R.; Kabeli, O.; Shmueli, Y.; Dos Santos Costa, A.; Fazel-Zarandi, M.; Sercu, T.; Candido, S.; Rives, A. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 2023, 379, 1123–1130.

[3] Heinzinger, M.; Weissenow, K.; Sanchez, J. G.; Henkel, A.; Steinegger, M.; Rost, B. ProstT5: Bilingual language model for protein sequence and structure. bioRxiv 2023, 2023–07.