Related Experiment Video
Updated: May 19, 2026
06:50
Computational Prediction of Amino Acid Preferences of Potentially Multispecific Peptide-Binding Domains Involved in Protein-Protein Interactions
Published on: January 26, 2024
Integrating Diffusion and Liquid AI Models for Predicting Peptide Affinity from mRNA Display Selections.
Colin M Leaf1, Pearl Qi2, Yash Pragnesh Gandhi3
1Department of Chemistry, University of Southern California, Los Angeles, California, 90089, USA.
Biorxiv : the Preprint Server for Biology
|May 18, 2026
Summary
Machine learning models predict peptide binding energies, expanding ligand discovery. This approach accurately identifies high-affinity peptide candidates generated through directed evolution and AI, accelerating the search for novel therapeutics.
Area of Science:
- Biochemistry and Molecular Biology
- Computational Biology and Bioinformatics
- Drug Discovery and Development
Background:
- In vitro selection and directed evolution technologies like mRNA display explore vast libraries to identify functional polypeptide ligands.
- Machine learning models, specifically Denoising Diffusion Implicit Models (DDIMs), trained on deep sequencing data can generate novel peptide sequences beyond experimental reach.
- Predicting peptide properties, such as binding free energies (ΔG°), is crucial for advancing ligand discovery but remains challenging.
Purpose of the Study:
- To apply machine learning methods for predicting binding free energies (ΔG°) of peptide ligands against the oncogenic protein Bcl-x L.
- To evaluate the accuracy of these predictions for both experimentally derived and DDIM-generated peptide sequences.
- To establish a unified computational framework for expanding experimental ligand discovery and predicting molecular properties of peptide candidates.
Main Methods:
- Trained a Closed-form Continuous (CfC) neural network using a dataset of 15,700 peptide ligands with known sequences and binding free energies (ΔG°).
- Utilized deep sequencing data from mRNA display libraries to train DDIM models for generating novel peptide sequences.
- Applied the trained CfC model to predict binding free energies for both experimental and DDIM-generated peptides.
Main Results:
- The CfC model accurately predicted the rank order and binding free energies (ΔG°) of peptide ligands within experimental error.
- Identified five DDIM-generated peptides exhibiting single-digit picomolar binding affinities.
- Demonstrated the model's capability to predict properties for both experimental and computationally generated peptide sequences.
Conclusions:
- Combining trained DDIM and CfC models provides a powerful, unified approach for in silico ligand discovery.
- This integrated strategy significantly expands the scope of identifying high-affinity peptide candidates.
- Highlights the utility of large quantitative datasets for accurate prediction of peptide molecular properties and therapeutic potential.
