The Ascher (Biosig) lab recently reported DDMut-PPI, a deep learning method for the accurate prediction of the binding affinity of protein-protein interactions (PPIs). DDMut-PPI builds on a previously published model called DDMut by the same group. DDMut-PPI predicts the effects of both single and multiple point mutations.
Computational Methods to Evaluate PPIs
Computational methods are essential for evaluating PPIs, providing a cost-effective alternative to experimental approaches. Techniques include molecular docking, molecular dynamics simulations, and machine learning models like DDMut-PPI. These methods help predict the strength and specificity of interactions between protein pairs, offering insights into potential therapeutic targets and the effects of mutations on these interactions.
Experimentally Determined PPIs in PDB
The Protein Data Bank (PDB) is a crucial resource for structural biology, containing over 180,000 experimentally determined protein structures. Approximately 15% of these structures represent protein-protein complexes. These structures, while could provide valuable data for training and validating computational models like DDMut-PPI, still represent a small portion of PPI available in proteome.
Physical Properties of PPIs
PPIs are characterized by several physical properties, including binding affinity, specificity, and the nature of the interacting surfaces. These interactions often involve hydrogen bonds, hydrophobic interactions, and van der Waals forces, contributing to the stability and specificity of the protein complexes. Understanding these properties is crucial for designing molecules that can modulate PPIs, offering potential therapeutic benefits.
Importance of Understanding PPIs for Drug Development
PPIs are crucial for understanding diseases because of their involvement in various cellular processes, making them an important focus of targeted therapeutic interventions. Accurate methods to predict the impact of mutations on the binding free energy of PPIs can be important for drug design, in particular for the development of Biologics. Monoclonal antibodies are designed to target specific PPIs, disrupting pathological interactions or enhancing beneficial ones. By accurately predicting the effects of mutations on PPIs, models like DDMut-PPI can aid in the design of more effective therapeutic agents, reducing the time and cost associated with drug development.
DDMut and DDMut-PPI
The parent model of DDMut-PPI, called DDMut, is based on Siamese neural networks with graph-based signatures. DDMut-PPI improves on this by using graph CNN focused on protein-protein interfaces, facilitated by embeddings from the ProtT5 protein language model.
To train DDMut-PPI, the SKEMPI 2.0 S4169 diverse dataset was used. This dataset comprises 3,268 binding affinity-reducing mutations and 901 affinity-enhancing mutations, spanning 319 unique protein complexes classified into 138 interaction types, such as protease–inhibitor, antibody–antigen, T-cell receptor–peptide, and so on. The dataset was augmented with an additional 4,169 hypothetical reverse mutations where each reverse mutation was paired with its corresponding forward mutation, resulting in the S8338 dataset.
Several blind test sets were used to validate the model, including the AB-Bind S645 dataset comprising 645 mutations across 32 antibody–antigen complexes, a collection of experimental mutational binding affinity data for the MDM2-p53 complex, deep mutational scanning datasets for the SPIKE–ACE2 complex, and more.
To evaluate DDMut-PPI for the prediction of the effects of multiple mutations, multiple point mutations datasets from the SKEMPI 2.0 database were used, including the SM1124 dataset (for double and triple point mutations) and the SM595 dataset (for >3 residue mutations).
Evaluation of DDMut-PPI showed that it achieves state-of-the-art performance with a Pearson correlation of up to 0.75 on single mutations, and up to 0.83 on multiple mutations.
Starting from a user-provided structure, the DDMut-PPI model offers several analysis types:
1️⃣ Single Mutation: For the prediction of the effect of single point mutations.
2️⃣ Mutation List: For a list of single point mutations.
3️⃣ Interface Analysis: For alanine scanning and site-saturation mutagenesis of interface residues.
4️⃣ Multiple Mutations: For combinatorial mutations at multiple sites as well as permutations of double and triple point mutations.
References
- Paper: DDMut-PPI: predicting effects of mutations on protein–protein interactions using graph-based deep learning
- Web server: https://biosig.lab.uq.edu.au/ddmut_ppi/
- API documentation
- DDMut-PPI Data Sets: https://biosig.lab.uq.edu.au/ddmut_ppi/datasets
- Moal, I. H., & Fernández-Recio, J. (2019). SKEMPI 2.0: An updated benchmark of changes in protein–protein binding energy, kinetics and thermodynamics upon mutation. Bioinformatics, 35(19), 4686-4693.
- Keskin, O., Tuncbag, N., & Gursoy, A. (2016). Predicting Protein-Protein Interactions from the Molecular to the Proteome Level. Chemical Reviews, 116(8), 4884–4909.
- Mak, A. B., et al. (2014). A lentiviral functional proteomics approach identifies chromatin remodeling complexes important for the induction of pluripotency. Molecular Cell Proteomics, 13(3), 392–403.
- Mullard, A. (2018). The antibody-drug conjugates boom. Nature Reviews Drug Discovery, 17(12), 803-804.
