data and model weights for MotifAE, a sparse autoencoder for unsupervised discovery of functional motifs from protein language model.

representative_2.3M_seq.csv contains representative proteins from Foldseek 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, MotifAE_k3_step_80000.pt, MotifAE_k4_step_80000.pt, and MotifAE_k5_step_80000.pt are checkpoints at 80,000 steps of these models. SAE was trained with reconstruction loss and L1 norm, MotifAE models were trained with an additional smoothness loss, k is smoothness loss window size. Three MotifAE models show similar performances on downstream tasks. 412pros_ddG_ML.csv contains the deep mutational scanning data of protein folding stability, which is use to train MotifAE-G. stability_associated_gate.pt contains selected features using MotifAE-G from the dataset.