SeaMoon: From protein language models to continuous structural heterogeneity

How proteins move and deform determines their interactions with the environment and is thus of the 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 sequences, without exploiting 3D structures. SeaMoon leverages protein language model (pLM) embeddings as input to a lightweight convolutional neural network. We assessed SeaMoon against $\sim$$\sim$1,000 collections of experimental conformations exhibiting diverse motions. It predicts at least one ground-truth motion with reasonable accuracy for 40% of the test proteins. SeaMoon captures motions inaccessible to normal mode analysis, an unsupervised physics-based method relying solely on 3D geometry, and generalizes to proteins without detectable sequence similarity to the training set. SeaMoon is easily retrainable with novel or updated pLMs.