“Using sparse autoencoders, researchers found that the latent space of protein language models decomposes into interpretable features corresponding to biological concepts like amino acids, structural motifs, and protein domains.”