MotifAE: Unsupervised Discovery of Functional Motifs from Protein Language Model

representative_2.3M_seq.csv contains representative proteins from structure-based clustering of Alphafold structure database. The ESM2-650M last layer embeddings of these proteins were used to train SAE and MotifAE.

SAE_step_80000.pt and MotifAE_step_80000.pt are checkpoints at 80,000 steps of both models. SAE was trained with reconstruction loss and L1 norm, MotifAE was trained with an additional local similarity loss.

412pros_ddG_ML.csv contains the deep mutational scanning data of protein folding stability, which is use to train MotifAE-G. 1404_stability_associated_features.pt were selected features using MotifAE-G.