With an accuracy of 0.96, a physical and chemical constraint graph neural network is used to predict protein-ligand interactions from sequences.

Editor | Radish Skin

In drug development, it is crucial to determine the binding affinity and functional effect of small molecule ligands on proteins. Current computational methods can predict these protein-ligand interaction properties, but without high-resolution protein structures, accuracy is often lost and functional effects cannot be predicted.

Researchers from Monash University and Griffith University have developed PSICHIC (PhySIcoCHhemICal graph neural network), a framework that combines physicochemical constraints to decode interaction fingerprints directly from sequence data. This enables PSICHIC to decode the mechanisms behind protein-ligand interactions, achieving state-of-the-art accuracy and interpretability.

Trained on the same protein-ligand pairs without structural data, PSICHIC performs on par with, or even exceeds, leading structure-based methods in binding affinity predictions.

PSICHIC’s interpretable fingerprints identify protein residues and ligand atoms involved in interactions and help reveal the selectivity determinants of protein-ligand interactions.

The research was titled "Physicochemical graph neural network for learning protein–ligand interaction fingerprints from sequence data" and was published in "Nature Machine Intelligence" on June 17, 2024.

Challenges in protein-ligand affinity prediction

In drug discovery, it is important to determine the binding affinity and functional effect of small molecule ligands on proteins, as the selective interaction of the ligand with a specific protein determines The expected effect of the drug.

However, although current computational methods are capable of predicting protein-ligand interaction properties, without high-resolution protein structures, prediction accuracy is often reduced, and there are also difficulties in predicting functional effects.

Although sequence-based methods are more advantageous in terms of cost and resources (e.g., no expensive experimental structure determination process is required), these methods usually face the problem of excessive degrees of freedom in pattern matching, which easily leads to overfitting and limited generalization. ization capabilities, thereby creating a performance gap with structure- or composite-based methods.

Physical Chemistry Graph Neural Network

A research team from Monash University and Griffith University developed PSICHIC (Physical Chemistry Graph Neural Network), a method to directly decode protein-ligands from sequence data following physical and chemical principles. Body interaction fingerprint method. Unlike previous sequence-based models, PSICHIC specifically incorporates physicochemical constraints to achieve state-of-the-art accuracy and interpretability.

As a 2D sequence-based method, PSICHIC generates and imposes these constraints on a 2D plot by applying a clustering algorithm, allowing PSICHIC to primarily adapt to the rational underlying patterns that determine protein-ligand interactions during training.

Performance Validation and Comparison

After training on the same protein-ligand pairs without structural data, PSICHIC rivals or even surpasses state-of-the-art structure-based and complex-based methods in binding affinity predictions they.

Experimental results on PDBBind v2016 and PDBBind v2020 data sets show that PSICHIC outperforms other sequence-based methods, such as TransCPI, MolTrans and DrugBAN, on multiple indicators.

Graphic: Summary of performance statistics for protein-ligand binding affinity predictions on the PDBBind v2016 and PDBBind v2020 benchmarks. (Source: paper)

Specifically, PSICHIC shows lower prediction error and higher correlation index, especially in terms of prediction accuracy and generalization ability. PSICHIC achieves an accuracy of up to 0.96 in functional effect prediction.

Furthermore, PSICHIC excels in the identification of binding sites and key ligand functional groups. In the analysis of multiple protein-ligand complex structures (such as PDB 6K1S and 6OXV), PSICHIC was able to accurately locate important binding residues and ligand functional groups, which validated its ability to directly decode proteins in sequence data - The ability of ligands to interact with each other. This capability is particularly reflected in its ability to predict protein-ligand binding sites and key residues from sequence data.

1. PSICHIC’s Interpretable Fingerprint

Illustration: Virtual screening using interactive fingerprints. (Source: paper)

Interestingly, PSICHIC’s interpretable fingerprints show that it obtains the ability to decode the underlying mechanism of protein-ligand interactions from sequence data alone, identifying binding site protein residues and involved ligand atoms Ability. This is true even when training only on sequence data with binding affinity labels and no interaction information.

Illustration: Selectivity analysis using interaction fingerprints. (Source: Paper) Researchers used PSICHIC to successfully screen a new adenosine A1 receptor agonist (Tanimoto similarity to the closest known A1R agonist is 0.2), and analyzed the differences between adenosine receptor subtypes. ligand selectivity.
Value
Protein-ligand interaction fingerprints describe the characteristics of specific interactions that occur between ligands and protein residues. Traditionally, these fingerprints are derived from 3D protein-ligand complexes, an expensive process that this paper shows is sensitive to structural resolution quality.
In contrast, PSICHIC utilizes only sequence data, providing a unique approach to obtaining interpretable interaction fingerprints. By incorporating constraints, PSICHIC exhibits emerging capabilities that enable it to reveal protein-ligand interaction mechanisms and efficiently predict interaction properties. PSYCHIC eliminates the need for 3D data, paving the way for robust learning on large-scale sequence databases.
As a proof of concept, the team demonstrated that PSICHIC can effectively screen drug candidates and perform selectivity analysis. PSICHIC requires only sequence data to run and has the potential to become a universally useful tool in drug discovery. Researchers expect it to play a role in de novo ligand design, into which PSICHIC's interpretable fingerprints can be integrated to optimize molecular structures.
Future Outlook
Currently, PSICHIC is limited to analyzing protein-ligand interactions of single proteins. Future plans include extending its analysis to protein complexes, such as GPCRs complexed with heterotrimeric G proteins, which could facilitate comprehensive studies of protein-ligand dynamics directly from sequence data.
In addition, PSICHIC’s powerful learning capabilities from sequence data paves the way for exploring complex interactions such as allosteric regulation, helping to understand how allosteric ligands regulate orthosteric ligands within protein targets.
The team has made their data, code and optimization models available to the wider scientific community. PSICHIC has proven its robustness and effectiveness in various application areas, has broad potential for future development, and is expected to have a significant impact on the field of virtual compound screening and the design of innovative small molecule therapeutics.
Paper link: https://www.nature.com/articles/s42256-024-00847-1
Related reports: https://phys.org/news/2024-06-ai-tool-rapid-effective-drug .html

The above is the detailed content of With an accuracy of 0.96, a physical and chemical constraint graph neural network is used to predict protein-ligand interactions from sequences.. For more information, please follow other related articles on the PHP Chinese website!