Virtual Screening with Gnina 1.0

Virtual screening—predicting which compounds within a specified compound library bind to a target molecule, typically a protein—is a fundamental task in the field of drug discovery. Doing virtual screening well provides tangible practical benefits, including reduced drug development costs, faster time to therapeutic viability, and fewer unforeseen side effects. As with most applied computational tasks, the algorithms currently used to perform virtual screening feature inherent tradeoffs between speed and accuracy. Furthermore, even theoretically rigorous, computationally intensive methods may fail to account for important effects relevant to whether a given compound will ultimately be usable as a drug. Here we investigate the virtual screening performance of the recently released Gnina molecular docking software, which uses deep convolutional networks to score protein-ligand structures. We find, on average, that Gnina outperforms conventional empirical scoring. The default scoring in Gnina outperforms the empirical AutoDock Vina scoring function on 89 of the 117 targets of the DUD-E and LIT-PCBA virtual screening benchmarks with a median 1% early enrichment factor that is more than twice that of Vina. However, we also find that issues of bias linger in these sets, even when not used directly to train models, and this bias obfuscates to what extent machine learning models are achieving their performance through a sophisticated interpretation of molecular interactions versus fitting to non-informative simplistic property distributions.

Virtual Screening with Gnina 1.0

Virtual screening - predicting which compounds within a specified compound library bind to a target molecule, typically a protein - is a fundamental task in the field of drug discovery. Doing virtual screening well provides tangible practical benefits, including reduced drug development costs, faster time to therapeutic viability, and fewer unforeseen side effects. As with most applied computational tasks, the algorithms currently used to perform virtual screening feature inherent tradeoffs between speed and accuracy. Furthermore, even theoretically rigorous, computationally intensive methods may fail to account for important effects relevant to whether a given compound will ultimately be usable as a drug. Here we investigate the virtual screening performance of the recently released Gnina molecular docking software, which uses deep convolutional networks to score protein-ligand structures. We find, on average, that Gnina outperforms conventional empirical scoring. The default scoring in Gnina outperforms the empirical AutoDock Vina scoring function on 89 of the 117 targets of the DUD-E and LIT-PCBA virtual screening benchmarks with a median 1% early enrichment factor that is more than twice that of Vina. However, we also find that issues of bias linger in these sets, even when not used directly to train models, and this bias obfuscates to what extent machine learning models are achieving their performance through a sophisticated interpretation of molecular interactions versus fitting to non-informative simplistic property distributions.

Can molecular dynamics simulations improve the structural accuracy and virtual screening performance of GPCR models?

The determination of G protein-coupled receptor (GPCR) structures at atomic resolution has improved understanding of cellular signaling and will accelerate the development of new drug candidates. However, experimental structures still remain unavailable for a majority of the GPCR family. GPCR structures and their interactions with ligands can also be modelled computationally, but such predictions have limited accuracy. In this work, we explored if molecular dynamics (MD) simulations could be used to refine the accuracy of in silico models of receptor-ligand complexes that were submitted to a community-wide assessment of GPCR structure prediction (GPCR Dock). Two simulation protocols were used to refine 30 models of the D3 dopamine receptor (D3R) in complex with an antagonist. Close to 60 μs of simulation time was generated and the resulting MD refined models were compared to a D3R crystal structure. In the MD simulations, the transmembrane helix region of the models generally drifted further away from the crystal structure conformation. However, MD refinement was able to improve the accuracy of the ligand binding mode and the second extracellular loop region. The best refinement protocol improved agreement with the experimentally observed ligand binding mode for a majority of the models. Receptor structures with improved virtual screening performance, which was assessed by molecular docking of ligands and decoys, could also be identified among the MD refined models. Application of weak restraints to the transmembrane helixes in the MD simulations further improved predictions of the ligand binding mode and second extracellular loop. These results provide guidelines for application of MD refinement in prediction of GPCR-ligand complexes and directions for further method development.

A Multi-Objective Approach for Drug Repurposing in Preeclampsia

Preeclampsia is a hypertensive disorder that occurs during pregnancy. It is a complex disease with unknown pathogenesis and the leading cause of fetal and maternal mortality during pregnancy. Using all drugs currently under clinical trial for preeclampsia, we extracted all their possible targets from the DrugBank and ChEMBL databases and labeled them as “targets”. The proteins labeled as “off-targets” were extracted in the same way but while taking all antihypertensive drugs which are inhibitors of ACE and/or angiotensin receptor antagonist as query molecules. Classification models were obtained for each of the 55 total proteins (45 targets and 10 off-targets) using the TPOT pipeline optimization tool. The average accuracy of the models in predicting the external dataset for targets and off-targets was 0.830 and 0.850, respectively. The combinations of models maximizing their virtual screening performance were explored by combining the desirability function and genetic algorithms. The virtual screening performance metrics for the best model were: the Boltzmann-Enhanced Discrimination of ROC (BEDROC)α=160.9 = 0.258, the Enrichment Factor (EF)1% = 31.55 and the Area Under the Accumulation Curve (AUAC) = 0.831. The most relevant targets for preeclampsia were: AR, VDR, SLC6A2, NOS3 and CHRM4, while ABCG2, ERBB2, CES1 and REN led to the most relevant off-targets. A virtual screening of the DrugBank database identified estradiol, estriol, vitamins E and D, lynestrenol, mifrepristone, simvastatin, ambroxol, and some antibiotics and antiparasitics as drugs with potential application in the treatment of preeclampsia.

Generating property-matched decoy molecules using deep learning

Motivation An essential step in the development of virtual screening methods is the use of established sets of actives and decoys for benchmarking and training. However, the decoy molecules in commonly used sets are biased meaning that methods often exploit these biases to separate actives and decoys, and do not necessarily learn to perform molecular recognition. This fundamental issue prevents generalization and hinders virtual screening method development. Results We have developed a deep learning method (DeepCoy) that generates decoys to a user’s preferred specification in order to remove such biases or construct sets with a defined bias. We validated DeepCoy using two established benchmarks, DUD-E and DEKOIS 2.0. For all 102 DUD-E targets and 80 of the 81 DEKOIS 2.0 targets, our generated decoy molecules more closely matched the active molecules’ physicochemical properties while introducing no discernible additional risk of false negatives. The DeepCoy decoys improved the Deviation from Optimal Embedding (DOE) score by an average of 81% and 66%, respectively, decreasing from 0.166 to 0.032 for DUD-E and from 0.109 to 0.038 for DEKOIS 2.0. Further, the generated decoys are harder to distinguish than the original decoy molecules via docking with Autodock Vina, with virtual screening performance falling from an AUC ROC of 0.70 to 0.63. Availability and implementation The code is available at https://github.com/oxpig/DeepCoy. Generated molecules can be downloaded from http://opig.stats.ox.ac.uk/resources. Supplementary information Supplementary data are available at Bioinformatics online.

Generating Property-Matched Decoy Molecules Using Deep Learning

An essential step in the development of virtual screening methods is the use of established sets of actives and decoys for benchmarking and training. However, the decoy molecules in commonly used sets are biased meaning that methods often exploit these biases to separate actives and decoys, rather than learning how to perform molecular recognition. This fundamental issue prevents generalisation and hinders virtual screening method development. We have developed a deep learning method (DeepCoy) that generates decoys to a user’s preferred specification in order to remove such biases or construct sets with a defined bias. We validated DeepCoy using two established benchmarks, DUD-E and DEKOIS 2.0. For all DUD-E targets and 80 of the 81 DEKOIS 2.0 targets, our generated decoy molecules more closely matched the active molecules’ physicochemical properties while introducing no discernible additional risk of false negatives. The DeepCoy decoys improved the Deviation from Optimal Embedding (DOE) score by an average of 81% and 66%, respectively, decreasing from 0.163 to 0.032 for DUD-E and from 0.109 to 0.038 for DEKOIS 2.0. Further, the generated decoys are harder to distinguish than the original decoy molecules via docking with Autodock Vina, with virtual screening performance falling from an AUC ROC of 0.71 to 0.63. The code is available at https://github.com/oxpig/DeepCoy. Generated molecules can be downloaded from http://opig.stats.ox.ac.uk/resources.

Improving structure-based virtual screening performance via learning from scoring function components

Scoring functions (SFs) based on complex machine learning (ML) algorithms have gradually emerged as a promising alternative to overcome the weaknesses of classical SFs. However, extensive efforts have been devoted to the development of SFs based on new protein–ligand interaction representations and advanced alternative ML algorithms instead of the energy components obtained by the decomposition of existing SFs. Here, we propose a new method named energy auxiliary terms learning (EATL), in which the scoring components are extracted and used as the input for the development of three levels of ML SFs including EATL SFs, docking-EATL SFs and comprehensive SFs with ascending VS performance. The EATL approach not only outperforms classical SFs for the absolute performance (ROC) and initial enrichment (BEDROC) but also yields comparable performance compared with other advanced ML-based methods on the diverse subset of Directory of Useful Decoys: Enhanced (DUD-E). The test on the relatively unbiased actives as decoys (AD) dataset also proved the effectiveness of EATL. Furthermore, the idea of learning from SF components to yield improved screening power can also be extended to other docking programs and SFs available.