Ribosomally synthesized and post-translationally modified peptides (RiPPs) comprise a structurally and functionally diverse group of natural products. Lasso peptides represent one of about 50 known molecular classes of RiPPs, which display a characteristic [1]rotaxane conformation formed by a lasso cyclase. This unique, threaded conformation endows lasso peptides with diverse biological activities and remarkable thermal and proteolytic stability. The prediction of lasso peptide properties, such as substrate compatibility with a particular lasso cyclase or desired biological activity, remains challenging due to limited experimental data and the intricate nature of the substrate fitness landscapes. Protein language models (PLMs) have demonstrated impressive performance in predicting protein structure and function. However, general-purpose PLMs perform poorly in lasso peptide-related predictive tasks. Therefore, there is a need to provide effective representations for lasso peptides to enable enhanced property prediction. In this study, we developed a lasso peptide-specific language model (LassoESM) by leveraging advances in pre-trained PLMs to aid the prediction of lasso peptide related properties and experimentally validate the model predictions. We demonstrate that LassoESM embeddings can accurately predict substrate compatibility for a lasso cyclase of interest when experimental data for model training was scarce. Using a deep learning framework incorporating cross-attention between lasso cyclase and substrate peptide embeddings, we identify non-cognate pairs of lasso cyclases and substrate peptides with predicted compatibility. We further show that LassoESM embeddings improve the prediction of RNA polymerase inhibitory activity, which represents a biological activity of several known lasso peptides. We anticipate that LassoESM and future iterations thereof will be instrumental for the rational design of lasso peptides with desired properties.
