LassoESM: A tailored language model for enhanced lasso peptide property prediction

Table of Contents

• Quick Start
  • Getting started with this repo
  • Usage
• Repository Structure
  • code
  • example_notebook
  • data
• Reference

Quick Start

Getting started with this repo

git clone https://github.com/xuenanmi/LassoESM.git
cd LassoESM
conda env create -f environment.yml
conda activate lassoesm

Usage

The script code/LassoESM/get_embeddings_LassoESM.py can be used to extract embeddings for peptide sequences from LassoESM/PeptideESM/VanillaESM

Repository Structure

code

Source code for model training and downstream tasks:

• LassoESM pretraining:

  • LassoESM_pretraining.py: Pretraining Lasso Peptide-Specific Language Model
  • get_embeddings_LassoESM: extract embeddings for peptide variants in dataset from LassoESM/PeptideESM/VanillaESM, then feed them into various downstream classification models

• Downstream Task 1: Substrate Tolerance Prediction

  • hyperparameter_optimization_ML_FusA.py: grid-search for hyperparameters of downstream classification models
  • downstream_models_performance_diff_4embs.py: compare downstream model performance with different embeddings (VanillaESM, PeptideESM, LassoESM)
  • diff_training_size.py: evaulate downstream model performance using different training size
  • cal_uncertainty.py: explore uncertainty of classification model output

• Downstream Task 2: Cyclase-Peptide Pair Prediction

  • generate_negative_cyclase_peptide_pairs.py: generate the synthetic cyclase-peptide pairs (negative samples)
  • predict_cyclase_peptide_pairs: a general model (MLP) to predict cyclase(embedded by VanillaESM)-peptide(embedded by LassoESM) pairs
  • predict_cyclase_peptide_pairs_with_CrossAttention: add a cross-attention layer, where lasso peptide embeddings reweight its corresponding cyclase embeddings
  • predict_non_natural_cyclase_peptide_pairs.py: use the trained cyclase-peptide pair prediction model to assess the compatibility of FusC with other predicted naturally occuring lasso peptides

• Downstream Task 3: Antimicrobial Activity Prediction

  • get_embeddings_for_Ubonodin.py: extracting embeddings for Ubonodin variant sequences from LassoESM/PeptideESM/VanillaESM models
  • antimicrobial_activity_prediction_for_Ubonodin.py: predict the antimicrobial activity of Ubonodin variant sequences
  • get_embeddings_for_Klebsidin.py: extracting embeddings for Klebsidin variant sequences from LassoESM/PeptideESM/VanillaESM models
  • antimicrobial_activity_prediction_for_Klebsidin.py: predict the antimicrobial activity of Klebsidin variant sequences

---

example_notebook

Example Jupyter notebooks for key analysis:

• extract_embeddings.ipynb: Demonstrates embedding extraction workflow.
• predict_substrate_tolerance.ipynb: Substrate tolerance classification pipeline.
• predict_cyclase_peptide_pairs.ipynb: Predict cyclase-peptide compatibility.
• predict_antimicrobial_activity.ipynb: Evaluate peptide antimicrobial activity.

---

data

Dataset used in the paper

• data_for_LassoESM_training
• data_for_substrate_tolerance_prediction
• data_for_cyclase_peptide_pair_prediction
• data_for_antimicrobial_activity_prediction

---

Reference

Mi, X., Barrett, S.E., Mitchell, D.A. et al. LassoESM a tailored language model for enhanced lasso peptide property prediction. Nat Commun 16, 8545 (2025). https://doi.org/10.1038/s41467-025-63412-3