The success rate exceeds that of the RoseTTAFold series, using sequence information to directly predict protein-ligand complex structures.

Editor | Radish Skin

Protein is a well-established tool in the body's fight against pathogens and is used to narrow down the range of potential treatments for experimental testing. High-quality protein structure is required, and proteins are often viewed as fully or partially rigid.

Here, researchers at Freie Universität Berlin have developed an artificial intelligence system that can predict fully flexible all-atom structures of protein-ligand complexes directly from sequence information.

Although classical docking methods are still superior, this also depends on the crystal structure of the target protein. In addition to predicting flexible all-atom structures, the prediction confidence metric (plDDT) can be used to select accurate predictions and differentiate between strong and weak binders.

The study is titled "Structure prediction of protein-ligand complexes from sequence information with Umol" and was published in "Nature Communications on May 28, 2024 》.

#Protein-protein target contacts are important issues in evaluating new drugs and repurposing known ones. Existing contact methods have limitations: they require high-quality protein structures; it is difficult to determine accurate contact postures; they are mostly based on binding ability (affinity) evaluation, which is difficult to reflect other factors such as structural stability. However, existing contact methods are limited by the need for high-quality protein structures, accurate contact poses, and multi-based affinity assessment. Therefore, the exploration of new ligands is limited by a combined approach of protein assembly and structure evaluation.

Although machine learning has been applied in this field, its performance on known target areas has not surpassed the classic methods based on scoring functions. Moreover, the predicted protein structure is often not suitable for direct use in ligand docking.

In addition, if the structures in the evaluation set are divided based on release time rather than similarity, bias will be introduced, especially when facing receptor structures not seen in training, the performance will be halved.

Protein flexibility is crucial for reaching the binding state and successful docking. Although RoseTTAFold All-Atom can bind ligands when predicting proteins, its success rate on the PoseBusters test set is only 42%, and it is The behavior of unseen proteins is unknown, indicating that the challenge of protein-ligand complex structure prediction has not yet been fully resolved.

A team at Freie Universität Berlin has developed an AI method that can predict the structure of protein-ligand complexes based on sequence information by extending EvoFormer in AlphaFold2. This network is similar to RFAA except that 3D trajectories are not included and template structures or additional crystallographic ligand data are used as input or during training.

Illustration: Umol overview. (Source: Paper)

Starting from a protein sequence, alternative protein target sites (pockets), and ligands SMILES creates a multiple sequence alignment (MSA) and bond matrix. From this, features are generated within the network and 3D structures are generated. Since no structural information is required to generate the final protein-ligand complex structure, there are no restrictions on protein or ligand flexibility.

Umol achieves a higher success rate (SR, ligand RMSD ≤ 2 Å) when including pocket information on the PoseBusters test set compared to the closest RoseTTAFold All-Atom and NeuralPlexer1, 45 respectively %, 42%, and 24%, making it the best performing method in protein-ligand structure prediction.

Illustration: Prediction accuracy. (Source: paper)

When pocket information is removed from Umol and template information is removed from RFAA, the SR drops to 18% and 8% respectively. When using DiffDock with AF prediction, the accuracy is 21% but depends on highly accurate interface prediction (pocket RMSD

Many ligand poses slightly above the 2 Å success threshold may be comparable, suggesting that a more flexible scoring system may be needed. Umol's success rate exceeds AutoDock Vina at the 2.35 Å threshold. Even small alignment errors can become problematic when native protein structures are not used for scoring.

Cofolded protein-ligand complexes have the potential to accelerate drug repositioning. In particular, the researchers found that the predicted lDDT of the ligand (plDDT) can be used to select accurate docking poses, while the pIDDT of the protein pocket is suitable for selecting accurate interfaces.

Illustration: Confidence metrics and accuracy. (Source: Paper)

ligand plDDT also separated high-affinity ligands from low-affinity ligands, suggesting that some of the predictions for Umol and Umol-pocket uncertainty may be weak binders. This further demonstrates the capabilities of Umol and highlights that important aspects of protein-ligand interactions appear to be understood.

Illustration: BindingDB prediction. (Source: paper)

Despite the 18% accuracy without pocket information, the network can still differentiate between strong and weak binders to a certain extent. This is particularly useful for annotating unknown complexes, and the team presented 336 protein-ligand structures with very high confidence (ligand plDDT>85). It should be noted that although these structures appear reasonable and their L-plDDT scores are high, they still need to be verified experimentally.

Illustration: Using Umol-pocket to analyze the relationship between predicted different features and ligand RMSD (LRMSD) on the PoseBusters test set (n=428). (Source: Paper)

The researchers did not find a clear relationship between the model's predictive performance and "different features associated with the same protein or ligand."

Illustration: The 5 most difficult structures. (Source: paper)

However, Umol-pocket was accurate in 3 out of 5 cases where other methods were difficult to predict. By inverting the trained network, new ligand-binding proteins or protein-binding ligands can be designed. Another option is to use transfer learning to create a generative diffusion model for the same purpose. In this case, the ligand or protein plDDT can be maximized in an attempt to create a high-affinity binder.

The current version of PDBbind contains data processed from the PDB in 2019. Since then, additional protein-ligand complexes have been submitted, suggesting that greater accuracy may be achievable.

However, it is currently unclear what precision is required to obtain meaningful protein-ligand docking results. The high accuracy of protein structure prediction is not achievable in tasks involving other molecules, such as small molecules or RNA.

Without protein co-evolutionary information, the accuracy of structure predictions rapidly decreases. Since there are no similar sources of information for small molecules or RNA, one has to rely on atomic representations.

Table: Success rate (percentage of ligands with RMSD≤2Å) on the PoseBuster benchmark set divided by sequence identity (seqid) for the PDBBind 2020 version. (Source: Paper)

# Researchers believe that pocket information is very effective, and without pocket information, deep learning methods seem prone to overfitting. This finding further corroborates the observation that although many molecules in the PoseBusters test set contain highly similar analogues in the training data set, this similarity does not correlate with model success.

Illustration: Some tests. (Source: Paper)

The same degree of overfitting is not observed for structure-based docking methods such as Vina or Gold. This is expected since they are based on atomic scoring functions and therefore do not rely on protein homology to the same extent.

The deep learning method has significantly higher performance on the training set, indicating that protein homology plays an important role in protein-ligand docking. The performance of RFAA on the test set is higher than that on the training set, which indicates possible data leakage between the training and test sets.

In summary, we are still a long way from fully grasping the complexity of protein-ligand interactions, but using deep learning to predict the structure of the entire complex may bring scientists closer to a solution.

Umol: https://github.com/patrickbryant1/Umol

Paper link: https://www.nature.com/articles/s41467 -024-48837-6

The above is the detailed content of The success rate exceeds that of the RoseTTAFold series, using sequence information to directly predict protein-ligand complex structures.. For more information, please follow other related articles on the PHP Chinese website!