Extracting prime protein targets as possible drug candidates: machine learning evaluation.

Extracting "high ranking" or "prime protein targets" (PPTs) as potent MRSA drug candidates from a given set of ligands is a key challenge in efficient molecular docking. This study combines protein-versus-ligand matching molecular docking (MD) data extracted from 10 independent molecular docking (MD) evaluations - ADFR, DOCK, Gemdock, Ledock, Plants, Psovina, Quickvina2, smina, vina, and vinaxb to identify top MRSA drug candidates. Twenty-nine active protein targets (APT) from the enhanced DUD-E repository ( http://DUD-E.decoys.org ) are matched against 1040 ligands using "forward modeling" machine learning for initial "data mining and modeling" (DDM) to extract PPTs and the corresponding high affinity ligands (HALs). K-means clustering (KMC) is then performed on 400 ligands matched against 29 PTs, with each cluster accommodating HALs, and the corresponding PPTs. Performance of KMC is then validated against randomly chosen head, tail, and middle active ligands (ALs). KMC outcomes have been validated against two other clustering methods, namely, Gaussian mixture model (GMM) and density based spatial clustering of applications with noise (DBSCAN). While GMM shows similar results as with KMC, DBSCAN has failed to yield more than one cluster and handle the noise (outliers), thus affirming the choice of KMC or GMM. Databases obtained from ADFR to mine PPTs are then ranked according to the number of the corresponding HAL-PPT combinations (HPC) inside the derived clusters, an approach called "reverse modeling" (RM). From the set of 29 PTs studied, RM predicts high fidelity of 5 PPTs (17%) that bind with 76 out of 400, i.e., 19% ligands leading to a prediction of next-generation MRSA drug candidates: PPT2 (average HPC is 41.1%) is the top choice, followed by PPT14 (average HPC 25.46%), and then PPT15 (average HPC 23.12%). This algorithm can be generically implemented irrespective of pathogenic forms and is particularly effective for sparse data.

Evaluation of consensus scoring methods for AutoDock Vina, smina and idock.

We investigated the application of consensus scoring using the freely available and open source structure-based virtual screening docking programs AutoDock Vina, smina and idock. These individual programs and several simple consensus scoring methods were tested for their ability to identify hits against 20 DUD-E benchmark targets using the AUC and EF1 metrics. We found that all of the consensus scoring methods, however normalized, fared worse, on average, than simply using the output from a single program, smina. Additionally, the effect of a significant increase in the run time of all three programs was tested to find if a longer run time yielded improved results. Our results indicated that a longer run time than the default had little impact on the performance of these three programs or on consensus scoring methods based on their output. Thus, we have found that using the smina program alone at default settings is the best approach for researchers that do not have access to a suite of commercial docking software packages.

Improving the Virtual Screening Ability of Target-Specific Scoring Functions Using Deep Learning Methods.

Scoring functions play an important role in structure-based virtual screening. It has been widely accepted that target-specific scoring functions (TSSFs) may achieve better performance compared with universal scoring functions in actual drug research and development processes. A method that can effectively construct TSSFs will be of great value to drug design and discovery. In this work, we proposed a deep learning-based model named DeepScore to achieve this goal. DeepScore adopted the form of PMF scoring function to calculate protein-ligand binding affinity. However, different from PMF scoring function, in DeepScore, the score for each protein-ligand atom pair was calculated using a feedforward neural network. Our model significantly outperformed Glide Gscore on validation data set DUD-E. The average ROC-AUC on 102 targets was 0.98. We also combined Gscore and DeepScore together using a consensus method and put forward a consensus model named DeepScoreCS. The comparison results showed that DeepScore outperformed other machine learning-based TSSFs building methods. Furthermore, we presented a strategy to visualize the prediction of DeepScore. All of these results clearly demonstrated that DeepScore would be a useful model in constructing TSSFs and represented a novel way incorporating deep learning and drug design.

Interaction prediction in structure-based virtual screening using deep learning.

We introduce a deep learning architecture for structure-based virtual screening that generates fixed-sized fingerprints of proteins and small molecules by applying learnable atom convolution and softmax operations to each molecule separately. These fingerprints are further non-linearly transformed, their inner product is calculated and used to predict the binding potential. Moreover, we show that widely used benchmark datasets may be insufficient for testing structure-based virtual screening methods that utilize machine learning. Therefore, we introduce a new benchmark dataset, which we constructed based on DUD-E, MUV and PDBBind databases.

Efficient conformational sampling and weak scoring in docking programs? Strategy of the wisdom of crowds.

BACKGROUND: In drug design, an efficient structure-based optimization of a ligand needs the precise knowledge of the protein-ligand interactions. In the absence of experimental information, docking programs are necessary for ligand positioning, and the choice of a reliable program is essential for the success of such an optimization. The performances of four popular docking programs, Gold, Glide, Surflex and FlexX, were investigated using 100 crystal structures of complexes taken from the Directory of Useful Decoys-Enhanced database. RESULTS: The ligand conformational sampling was rather efficient, with a correct pose found for a maximum of 84 complexes, obtained by Surflex. However, the ranking of the correct poses was not as efficient, with a maximum of 68 top-rank or 75 top-4 rank correct poses given by Glidescore. No relationship was found between either the sampling or the scoring performance of the four programs and the properties of either the targets or the small molecules, except for the number of ligand rotatable bonds. As well, no exploitable relationship was found between each program performance in docking and in virtual screening; a wrong top-rank pose may obtain a good score that allows it to be ranked among the most active compounds and vice versa. Also, to improve the results of docking, the strengths of the programs were combined either by using a rescoring procedure or the United Subset Consensus (USC). Oddly, positioning with Surflex and rescoring with Glidescore did not improve the results. However, USC based on docking allowed us to obtain a correct pose in the top-4 rank for 87 complexes. Finally, nine complexes were scrutinized, because a correct pose was found by at least one program but poorly ranked by all four programs. Contrarily to what was expected, except for one case, this was not due to weaknesses of the scoring functions. CONCLUSIONS: We conclude that the scoring functions should be improved to detect the correct poses, but sometimes their failure may be due to other varied considerations. To increase the chances of success, we recommend to use several programs and combine their results. Graphical abstract Summary of the results obtained by semi-rigid docking of crystallographic ligands. The docking was done on 100 protein-ligand X-ray structures, taken from the DUD-E database, and using four programs, Glide, Gold, Surflex and FlexX. Based on the docking results, we applied our United Subset Consensus method (USC), for which only the top4-rank poses are relevant. The number of complexes for which the best pose is correct, is represented by the gray boxes, the blue and red boxes correspond to the number of complexes with a correct pose ranked as the top 1 or within the top 4. A pose is considered correct when its root-mean-square deviation from the crystal structure is less than 2 Å.

vSDC: a method to improve early recognition in virtual screening when limited experimental resources are available.

BACKGROUND: In drug design, one may be confronted to the problem of finding hits for targets for which no small inhibiting molecules are known and only low-throughput experiments are available (like ITC or NMR studies), two common difficulties encountered in a typical academic setting. Using a virtual screening strategy like docking can alleviate some of the problems and save a considerable amount of time by selecting only top-ranking molecules, but only if the method is very efficient, i.e. when a good proportion of actives are found in the 1-10 % best ranked molecules. RESULTS: The use of several programs (in our study, Gold, Surflex, FlexX and Glide were considered) shows a divergence of the results, which presents a difficulty in guiding the experiments. To overcome this divergence and increase the yield of the virtual screening, we created the standard deviation consensus (SDC) and variable SDC (vSDC) methods, consisting of the intersection of molecule sets from several virtual screening programs, based on the standard deviations of their ranking distributions. CONCLUSIONS: SDC allowed us to find hits for two new protein targets by testing only 9 and 11 small molecules from a chemical library of circa 15,000 compounds. Furthermore, vSDC, when applied to the 102 proteins of the DUD-E benchmarking database, succeeded in finding more hits than any of the four isolated programs for 13-60 % of the targets. In addition, when only 10 molecules of each of the 102 chemical libraries were considered, vSDC performed better in the number of hits found, with an improvement of 6-24 % over the 10 best-ranked molecules given by the individual docking programs.Graphical abstractIn drug design, for a given target and a given chemical library, the results obtained with different virtual screening programs are divergent. So how to rationally guide the experimental tests, especially when only a few number of experiments can be made? The variable Standard Deviation Consensus (vSDC) method was developed to answer this issue. Left panel the vSDC principle consists of intersecting molecule sets, chosen on the basis of the standard deviations of their ranking distributions, obtained from various virtual screening programs. In this study Glide, Gold, FlexX and Surflex were used and tested on the 102 targets of the DUD-E database. Right panel Comparison of the average percentage of hits found with vSDC and each of the four programs, when only 10 molecules from each of the 102 chemical libraries of the DUD-E database were considered. On average, vSDC was capable of finding 38 % of the findable hits, against 34 % for Glide, 32 % for Gold, 16 % for FlexX and 14 % for Surflex, showing that with vSDC, it was possible to overcome the unpredictability of the virtual screening results and to improve them.

Benchmark of four popular virtual screening programs: construction of the active/decoy dataset remains a major determinant of measured performance.

BACKGROUND: In a structure-based virtual screening, the choice of the docking program is essential for the success of a hit identification. Benchmarks are meant to help in guiding this choice, especially when undertaken on a large variety of protein targets. Here, the performance of four popular virtual screening programs, Gold, Glide, Surflex and FlexX, is compared using the Directory of Useful Decoys-Enhanced database (DUD-E), which includes 102 targets with an average of 224 ligands per target and 50 decoys per ligand, generated to avoid biases in the benchmarking. Then, a relationship between these program performances and the properties of the targets or the small molecules was investigated. RESULTS: The comparison was based on two metrics, with three different parameters each. The BEDROC scores with α = 80.5, indicated that, on the overall database, Glide succeeded (score > 0.5) for 30 targets, Gold for 27, FlexX for 14 and Surflex for 11. The performance did not depend on the hydrophobicity nor the openness of the protein cavities, neither on the families to which the proteins belong. However, despite the care in the construction of the DUD-E database, the small differences that remain between the actives and the decoys likely explain the successes of Gold, Surflex and FlexX. Moreover, the similarity between the actives of a target and its crystal structure ligand seems to be at the basis of the good performance of Glide. When all targets with significant biases are removed from the benchmarking, a subset of 47 targets remains, for which Glide succeeded for only 5 targets, Gold for 4 and FlexX and Surflex for 2. CONCLUSION: The performance dramatic drop of all four programs when the biases are removed shows that we should beware of virtual screening benchmarks, because good performances may be due to wrong reasons. Therefore, benchmarking would hardly provide guidelines for virtual screening experiments, despite the tendency that is maintained, i.e., Glide and Gold display better performance than FlexX and Surflex. We recommend to always use several programs and combine their results. Graphical AbstractSummary of the results obtained by virtual screening with the four programs, Glide, Gold, Surflex and FlexX, on the 102 targets of the DUD-E database. The percentage of targets with successful results, i.e., with BDEROC(α = 80.5) > 0.5, when the entire database is considered are in Blue, and when targets with biased chemical libraries are removed are in Red.