Beta atomic contacts: identifying critical specific contacts in protein binding interfaces

Artificial intelligence in the prediction of protein-ligand interactions: recent advances and future directions

New drug production, from target identification to marketing approval, takes over 12 years and can cost around $2.6 billion. Furthermore, the COVID-19 pandemic has unveiled the urgent need for more powerful computational methods for drug discovery. Here, we review the computational approaches to predicting protein-ligand interactions in the context of drug discovery, focusing on methods using artificial intelligence (AI). We begin with a brief introduction to proteins (targets), ligands (e.g. drugs) and their interactions for nonexperts. Next, we review databases that are commonly used in the domain of protein-ligand interactions. Finally, we survey and analyze the machine learning (ML) approaches implemented to predict protein-ligand binding sites, ligand-binding affinity and binding pose (conformation) including both classical ML algorithms and recent deep learning methods. After exploring the correlation between these three aspects of protein-ligand interaction, it has been proposed that they should be studied in unison. We anticipate that our review will aid exploration and development of more accurate ML-based prediction strategies for studying protein-ligand interactions.

Co-Occurring Atomic Contacts for the Characterization of Protein Binding Hot Spots

A binding hot spot is a small area at a protein-protein interface that can make significant contribution to binding free energy. This work investigates the substantial contribution made by some special co-occurring atomic contacts at a binding hot spot. A co-occurring atomic contact is a pair of atomic contacts that are close to each other with no more than three covalent-bond steps. We found that two kinds of co-occurring atomic contacts can play an important part in the accurate prediction of binding hot spot residues. One is the co-occurrence of two nearby hydrogen bonds. For example, mutations of any residue in a hydrogen bond network consisting of multiple co-occurring hydrogen bonds could disrupt the interaction considerably. The other kind of co-occurring atomic contact is the co-occurrence of a hydrophobic carbon contact and a contact between a hydrophobic carbon atom and a π ring. In fact, this co-occurrence signifies the collective effect of hydrophobic contacts. We also found that the B-factor measurements of several specific groups of amino acids are useful for the prediction of hot spots. Taking the B-factor, individual atomic contacts and the co-occurring contacts as features, we developed a new prediction method and thoroughly assessed its performance via cross-validation and independent dataset test. The results show that our method achieves higher prediction performance than well-known methods such as Robetta, FoldX and Hotpoint. We conclude that these contact descriptors, in particular the novel co-occurring atomic contacts, can be used to facilitate accurate and interpretable characterization of protein binding hot spots.

A structural dissection of large protein-protein crystal packing contacts

With the rapid increase in crystal structures of protein-protein complexes deposited in the Protein Data Bank (PDB), more and more crystal contacts have been shown to have similar or even larger interface areas than biological interfaces. However, little attention has been paid to these large crystal packing contacts and their structural principles remain unknown. To address this issue, we used a comparative feature analysis to analyze the geometric and physicochemical properties of large crystal packing contacts by comparing two types of specific protein-protein interactions (PPIs), weak transient complexes and permanent homodimers. Our results show that although large crystal packing contacts have a similar interface area and contact size as permanent homodimers, they tend to be more planar, loosely packed and less hydrophobic than permanent homodimers and cannot form a central core region that is fully buried during interaction. However, the properties of large crystal packing contacts, except for the interface area and contact size, more closely resemble those of weak transient complexes. The large overlap between biological and large crystal packing contacts indicates that interface properties are not efficient indicators for classification of biological interfaces from large crystal packing contacts and finding other specific features urgently needed.

Use B-factor related features for accurate classification between protein binding interfaces and crystal packing contacts

Background

Distinction between true protein interactions and crystal packing contacts is important for structural bioinformatics studies to respond to the need of accurate classification of the rapidly increasing protein structures. There are many unannotated crystal contacts and there also exist false annotations in this rapidly expanding volume of data. Previous tools have been proposed to address this problem. However, challenging issues still remain, such as low performance when the training and test data contain mixed interfaces having diverse sizes of contact areas.

Methods and results

B factor is a measure to quantify the vibrational motion of an atom, a more relevant feature than interface size to characterize protein binding. We propose to use three features related to B factor for the classification between biological interfaces and crystal packing contacts. The first feature is the sum of the normalized B factors of the interfacial atoms in the contact area, the second is the average of the interfacial B factor per residue in the chain, and the third is the average number of interfacial atoms with a negative normalized B factor per residue in the chain. We investigate the distribution properties of these basic features and a compound feature on four datasets of biological binding and crystal packing, and on a protein binding-only dataset with known binding affinity. We also compare the cross-dataset classification performance of these features with existing methods and with a widely-used and the most effective feature interface area. The results demonstrate that our features outperform the interface area approach and the existing prediction methods remarkably for many tests on all of these datasets.

Conclusions

The proposed B factor related features are more effective than interface area to distinguish crystal packing from biological binding interfaces. Our computational methods have a potential for large-scale and accurate identification of biological interactions from the experimentally determined structural data stored at PDB which may have diverse interface sizes.

Integrating water exclusion theory into β contacts to predict binding free energy changes and binding hot spots

BACKGROUND

Binding free energy and binding hot spots at protein-protein interfaces are two important research areas for understanding protein interactions. Computational methods have been developed previously for accurate prediction of binding free energy change upon mutation for interfacial residues. However, a large number of interrupted and unimportant atomic contacts are used in the training phase which caused accuracy loss.

RESULTS

This work proposes a new method, βACVASA, to predict the change of binding free energy after alanine mutations. βACVASA integrates accessible surface area (ASA) and our newly defined β contacts together into an atomic contact vector (ACV). A β contact between two atoms is a direct contact without being interrupted by any other atom between them. A β contact's potential contribution to protein binding is also supposed to be inversely proportional to its ASA to follow the water exclusion hypothesis of binding hot spots. Tested on a dataset of 396 alanine mutations, our method is found to be superior in classification performance to many other methods, including Robetta, FoldX, HotPOINT, an ACV method of β contacts without ASA integration, and ACVASA methods (similar to βACVASA but based on distance-cutoff contacts). Based on our data analysis and results, we can draw conclusions that: (i) our method is powerful in the prediction of binding free energy change after alanine mutation; (ii) β contacts are better than distance-cutoff contacts for modeling the well-organized protein-binding interfaces; (iii) β contacts usually are only a small fraction number of the distance-based contacts; and (iv) water exclusion is a necessary condition for a residue to become a binding hot spot.

CONCLUSIONS

βACVASA is designed using the advantages of both β contacts and water exclusion. It is an excellent tool to predict binding free energy changes and binding hot spots after alanine mutation.

Binding affinity prediction for protein-ligand complexes based on β contacts and B factor

Accurate determination of protein-ligand binding affinity is a fundamental problem in biochemistry useful for many applications including drug design and protein-ligand docking. A number of scoring functions have been proposed for the prediction of protein-ligand binding affinity. However, accurate prediction is still a challenging problem because poor performance is often seen in the evaluation under the leave-one-cluster-out cross-validation (LCOCV). We introduce a new scoring function named B2BScore to improve the prediction performance. B2BScore integrates two physicochemical properties for protein-ligand binding affinity prediction. One is the property of β contacts. A β contact between two atoms requires no other atoms to interrupt the atomic contact and assumes that the two atoms should have enough direct contact area. The other is the property of B factor to capture the atomic mobility in the dynamic protein-ligand binding process. Tested on the PDBBind2009 data set, B2BScore shows superior prediction performance to existing methods on independent test data as well as under the LCOCV evaluation framework. In particular, B2BScore achieves a significant LCOCV improvement across 26 protein clusters-a big increase of the averaged Pearson's correlation coefficients from 0.418 to 0.518 and a significant decrease of standard deviation of the coefficients from 0.352 to 0.196. We also identified several important and intuitive contact descriptors of protein-ligand binding through the random forest learning in B2BScore. Some of these descriptors are closely related to contacts between carbon atoms without covalent-bond oxygen/nitrogen, preferred contacts of metal ions, interfacial backbone atoms from proteins, or π rings. Some others are negative descriptors relating to those contacts with nitrogen atoms without covalent-bond hydrogens or nonpreferred contacts of metal ions. These descriptors can be directly used to guide protein-ligand docking.