Random Forest Refinement of Pairwise Potentials for Protein–Ligand Decoy Detection

Systematic Improvement of the Performance of Machine Learning Scoring Functions by Incorporating Features of Protein-Bound Water Molecules

Water molecules at the ligand-protein interfaces play crucial roles in the binding of the ligands, but the behavior of protein-bound water is largely ignored in many currently used machine learning (ML)-based scoring functions (SFs). In an attempt to improve the prediction performance of existing ML-based SFs, we estimated the water distribution with a HydraMap (HM) method and then incorporated the features extracted from protein-bound waters obtained in this way into three ML-based SFs: RF-Score, ECIF, and PLEC. It was found that a combination of HM-based features can consistently improve the performance of all three SFs, including their scoring, ranking, and docking power. HydraMap-based features show consistently good performance with both crystal structures and docked structures, demonstrating their robustness for SFs. Overall, HM-based features, which are a statistical representation of hydration sites at protein-ligand interfaces, are expected to improve the prediction performance for diverse SFs.

The impact of cross-docked poses on performance of machine learning classifier for protein-ligand binding pose prediction

Structure-based drug design depends on the detailed knowledge of the three-dimensional (3D) structures of protein-ligand binding complexes, but accurate prediction of ligand-binding poses is still a major challenge for molecular docking due to deficiency of scoring functions (SFs) and ignorance of protein flexibility upon ligand binding. In this study, based on a cross-docking dataset dedicatedly constructed from the PDBbind database, we developed several XGBoost-trained classifiers to discriminate the near-native binding poses from decoys, and systematically assessed their performance with/without the involvement of the cross-docked poses in the training/test sets. The calculation results illustrate that using Extended Connectivity Interaction Features (ECIF), Vina energy terms and docking pose ranks as the features can achieve the best performance, according to the validation through the random splitting or refined-core splitting and the testing on the re-docked or cross-docked poses. Besides, it is found that, despite the significant decrease of the performance for the threefold clustered cross-validation, the inclusion of the Vina energy terms can effectively ensure the lower limit of the performance of the models and thus improve their generalization capability. Furthermore, our calculation results also highlight the importance of the incorporation of the cross-docked poses into the training of the SFs with wide application domain and high robustness for binding pose prediction. The source code and the newly-developed cross-docking datasets can be freely available at https://github.com/sc8668/ml_pose_prediction and https://zenodo.org/record/5525936 , respectively, under an open-source license. We believe that our study may provide valuable guidance for the development and assessment of new machine learning-based SFs (MLSFs) for the predictions of protein-ligand binding poses.

FFENCODER-PL: Pair Wise Energy Descriptors for Protein-Ligand Pose Selection

Scoring functions are the essential component in molecular docking methods. An accurate scoring function is expected to distinguish the native ligand pose from decoy poses. Our previous experience (Pei et al. J. Chem. Inf. Model. 2019, 59 (7), 3305-3315) proved that combining the random forest (RF) algorithm with knowledge-based potential functions can emphasize germane pair wise interactions and improve the performance of original knowledge-based potential functions on protein-ligand decoy detection. One of the most important potential function classes is the force field (FF) potential with one example being the Amber collection of FFs, which are widely available in the AMBER suite of simulation programs. However, for use in RF modeling studies, one needs pair wise energies that are hard to directly extract from Amber. To address this issue, FFENCODER-PL was constructed to calculate the pair wise energies based on the FF14SB and GAFF2 FFs in Amber. FFENCODER-PL was validated using 275 ligand and 21 protein-ligand structures. RF models were built by combining an RF classification algorithm with the pair wise energies calculated from FFENCODER-PL. CASF-2016 (Su et al. J. Chem. Inf. Model. 2019, 59, 895-913) was employed to test the performance of the resultant RF models, which outperformed 33 scoring functions on accuracy and native ranking tests. For the best decoy RMSD test, RF models give a best decoy with an RMSD of around 2 Å from the native pose after including the best decoy-decoy comparisons in the RF model. The relative importance of the RF algorithm and force field potentials was also tested with the results suggesting that both the RF algorithm and force field potentials are important and combining them is the only way to achieve high accuracy. Finally, FFENCODER-PL makes force field-based pair wise energies available for further development of machine learning-based scoring functions. The codes and data used in this paper can be found at https://github.com/JunPei000/Amber_protein_ligand_encoding.

RF-SVM: Identification of DNA-binding proteins based on comprehensive feature representation methods and support vector machine

Protein-DNA interactions play an important role in biological progress, such as DNA replication, repair, and modification processes. In order to have a better understanding of its functions, the one of the most important steps is the identification of DNA-binding proteins. We propose a DNA-binding protein predictor, namely, RF-SVM, which contains four types features, that is, pseudo amino acid composition (PseAAC), amino acid distribution (AAD), adjacent amino acid composition frequency (ACF) and Local-DPP. Random Forest algorithm is utilized for selecting top 174 features, which are established the predictor model with the support vector machine (SVM) on training dataset UniSwiss-Tr. Finally, RF-SVM method is compared with other existing methods on test dataset UniSwiss-Tst. The experimental results demonstrated that RF-SVM has accuracy of 84.25%. Meanwhile, we discover that the physicochemical properties of amino acids for OOBM770101(H), CIDH920104(H), MIYS990104(H), NISK860101(H), VINM940103(H), and SNEP660101(A) have contribution to predict DNA-binding proteins. The main code and datasets can gain in https://github.com/NiJianWei996/RF-SVM.

Improving Docking Power for Short Peptides Using Random Forest

In recent years, therapeutic peptides have gained a lot interest as demonstrated by the 60 peptides approved as drugs in major markets and 150+ peptides currently in clinical trials. However, while small molecule docking is routinely used in rational drug design efforts, docking peptides has proven challenging partly because docking scoring functions, developed and calibrated for small molecules, perform poorly for these molecules. Here, we present random forest classifiers trained to discriminate correctly docked peptides. We show that, for a testing set of 47 protein-peptide complexes, structurally dissimilar from the training set and previously used to benchmark AutoDock Vina's ability to dock short peptides, these random forest classifiers improve docking power from ∼25% for AutoDock scoring functions to an average of ∼70%. These results pave the way for peptide-docking success rates comparable to those of small molecule docking. To develop these classifiers, we compiled the ProptPep37_2021 data set, a curated, high-quality set of 322 crystallographic protein-peptides complexes annotated with structural similarity information. The data set also provides a collection of high-quality putative poses with a range of deviations from the crystallographic pose, providing correct and incorrect poses (i.e., decoys) of the peptide for each entry. The ProptPep37_2021 data set as well as the classifiers presented here are freely available.

RASPD+: Fast Protein-Ligand Binding Free Energy Prediction Using Simplified Physicochemical Features

The virtual screening of large numbers of compounds against target protein binding sites has become an integral component of drug discovery workflows. This screening is often done by computationally docking ligands into a protein binding site of interest, but this has the drawback of a large number of poses that must be evaluated to obtain accurate estimates of protein-ligand binding affinity. We here introduce a fast pre-filtering method for ligand prioritization that is based on a set of machine learning models and uses simple pose-invariant physicochemical descriptors of the ligands and the protein binding pocket. Our method, Rapid Screening with Physicochemical Descriptors + machine learning (RASPD+), is trained on PDBbind data and achieves a regression performance that is better than that of the original RASPD method and traditional scoring functions on a range of different test sets without the need for generating ligand poses. Additionally, we use RASPD+ to identify molecular features important for binding affinity and assess the ability of RASPD+ to enrich active molecules from decoys.

Pair Potentials as Machine Learning Features

Atom pairwise potential functions make up an essential part of many scoring functions for protein decoy detection. With the development of machine learning (ML) tools, there are multiple ways to combine potential functions to create novel ML models and methods. Potential function parameters can be easily extracted; however, it is usually hard to directly obtain the calculated atom pairwise energies from scoring functions. Amber, as one of the most popular suites of modeling programs, has an extensive history and library of force field potential functions. In this work, we directly used the force field parameters in ff94 and ff14SB from Amber and encoded them to calculate atom pairwise energies for different interactions. Two sets of structures (single amino acid set and a dipeptide set) were used to evaluate the performance of our encoded Amber potentials. From the comparison results between energy terms obtained from our encoding and Amber, we find energy difference within ±0.06 kcal/mol for all tested structures. Previously we have shown that the Random Forest (RF) model can help to emphasize more important atom pairwise interactions and ignore insignificant ones [Pei, J.; Zheng, Z.; Merz, K. M. J. Chem. Inf. Model. 2019, 59, 1919-1929]. Here, as an example of combining ML methods with traditional potential functions, we followed the same work flow to combine the RF models with force field potential functions from Amber. To determine the performance of our RF models with force field potential functions, 224 different protein native-decoy systems were used as our training and testing sets We find that the RF models with ff94 and ff14SB force field parameters outperformed all other scoring functions (RF models with KECSA2, RWplus, DFIRE, dDFIRE, and GOAP) considered in this work for native structure detection, and they performed similarly in detecting the best decoy. Through inclusion of best decoy to decoy comparisons in building our RF models, we were able to generate models that outperformed the score functions tested herein both on accuracy and best decoy detection, again showing the performance and flexibility of our RF models to tackle this problem. Finally, the importance of the RF algorithm and force field parameters were also tested and the comparison results suggest that both the RF algorithm and force field potentials are important with the ML scoring function achieving its best performance only by combining them together. All code and data used in this work are available at https://github.com/JunPei000/FFENCODER_for_Protein_Folding_Pose_Selection.

Refinement of pairwise potentials via logistic regression to score protein-protein interactions

Protein-protein interactions (PPIs) are ubiquitous and functionally of great importance in biological systems. Hence, the accurate prediction of PPIs by protein-protein docking and scoring tools is highly desirable in order to characterize their structure and biological function. Ab initio docking protocols are divided into the sampling of docking poses to produce at least one near-native structure, and then to evaluate the vast candidate structures by scoring. Concurrent development in both sampling and scoring is crucial for the deployment of protein-protein docking software. In the present work, we apply a machine learning model on pairwise potentials to refine the task of protein quaternary structure native structure detection among decoys. A decoy set was featurized using the Knowledge and Empirical Combined Scoring Algorithm 2 (KECSA2) pairwise potential. The highly unbalanced decoy set was then balanced using a comparison concept between native and decoy structures. The resultant comparison descriptors were used to train a logistic regression (LR) classifier. The LR model yielded the optimal performance for native detection among decoys compared with conventional scoring functions, while exhibiting lesser performance for the detection of low root mean square deviation decoy structures. Its deployment on an independent benchmark set confirms that the scoring function performs competitively relative to other scoring functions. The scripts used are available at https://github.com/TanemuraKiyoto/PPI-native-detection-via-LR.