This repository hosts software that utilises machine learning models to design novel mono and dual agonists peptides targeting human GCG and GLP-1 receptors. It features a codebase designed to facilitate peptide sequence optimization, leveraging a pre-trained ensemble of deep multi-task neural network models to tailor activity profiles according to specific requirements. Furthermore, the ensemble of convolutional multi-task neural network models can be trained on user-supplied peptide activity datasets, enabling the fine-tuning of ligand potency for dual molecular targets.

Please note that this code is using Python 3.8.

To begin, create virtualenv with conda:

conda create -n PeptideModels python=3.8

and activate it:

conda activate PeptideModels

Install pip:

conda install pip

and project requirements:

pip install -r requirements.txt

Then install the PeptideModels package by running: pip install -e . at the setup.py location level.

conda install -c conda-forge nb_conda

Firstly, an ensemble of multi-task convolutional neural networks is trained using peptide sequence data to learn the relationship between the peptide sequences and their potency levels at two human GPC receptors simultaneously. This trained ensemble is subsequently utilised to predict the potencies of unknown sequences against human GCG and GLP-1 receptors. Finally, ligand optimization is conducted to tailor the activity profile towards one of the three desired outcomes: selective potency at GCGR, selective potency at GLP-1R, or high-potency at both receptors.

Training:

• To train the ensemble of deep convolutional multi-task neural network models detailed in our paper for activity prediction across two molecular targets, using either the provided human GCG/GLP1 analogous training dataset or your own dataset, and please refer to: Tutorial-training_ensemble in the folder notebooks. To run the training, simply enter python train_main.py in the command line. Note that the ensemble training on the provided dataset of 125 peptide sequences takes around 2 h (processor 2.3 GHz 8-Core Intel Core i9).

Prediction:

• To use our pre-trained ensemble for predictions on your dataset of uncharacterised human GCGR/GLP-1R binding sequences, please refer to the: Tutorial-predicting_potencies in the notebooks folder. The running time for the provided datasets of 288 GCG and GLP-1 orthologs is approximately 10 minutes.

Optimization:

• To use our pre-trained ensemble for model-guided ligand optimization, please see:Tutorial-ligand_design in the notebooks directory. The sequence optimization process (given the parameters as in the tutorial) is estimated to take approximately 25 minutes.

Should you find this repository beneficial to your research, please cite our article:

Machine learning designs new GCGR/GLP-1R dual agonists with enhances biological potency

Puszkarska, A.M., Taddese, B., Revell, J. et al. Machine learning designs new GCGR/GLP-1R dual agonists with enhanced biological potency. Nat. Chem. (2024).

The folder supporting_experiments contains all necessary code and data for replicating the supporting findings reported in our paper. This includes the analysis of the training dataset, optimization and evaluation of machine learning models, analysis of the model-designed compounds, dimensionality reduction, statistical analysis, and more. For an in-depth guide on navigating the repository, please refer to README_SUPPORT.md.

Contact: Anna Puszkarska ( 📧 [email protected])